Previous Article in Journal
Assessment of 137Cs and 40K Transfer Factors in Croatian Agricultural Systems and Implications for Food Safety
 
 
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Review

Computer Vision for Low-Level Nuclear Waste Sorting: A Review

1
Department of Civil Engineering, McMaster University, Hamilton, ON L8S 4L7, Canada
2
Clean Energy Materials Sorting and Recycling, Laurentis Energy Partners, 44 Frid Street, Hamilton, ON L8P 4M4, Canada
*
Author to whom correspondence should be addressed.
Environments 2025, 12(8), 270; https://doi.org/10.3390/environments12080270
Submission received: 14 May 2025 / Revised: 15 July 2025 / Accepted: 29 July 2025 / Published: 5 August 2025

Abstract

Nuclear power is a low-emission and economically competitive energy source, yet the effective disposal and management of its associated radioactive waste can be challenging. Radioactive waste can be categorised as high-level waste (HLW), intermediate-level waste (ILW), and low-level waste (LLW). LLW primarily comprises materials contaminated during routine clean-up, such as mop heads, paper towels, and floor sweepings. While LLW is less radioactive compared to HLW and ILW, the management of LLW poses significant challenges due to the large volume that requires processing and disposal. The volume of LLW can be significantly reduced through sorting, which is typically performed manually in a labour-intensive way. Smart management techniques, such as computer vision (CV) and machine learning (ML), have great potential to help reduce the workload and human errors during LLW sorting. This paper provides a comprehensive review of previous research related to LLW sorting and a summative review of existing applications of CV in solid waste management. It also discusses state-of-the-art CV and ML algorithms and their potential for automating LLW sorting. This review lays a foundation for and helps facilitate the applications of CV and ML techniques in LLW sorting, paving the way for automated LLW sorting and sustainable LLW management.

1. Introduction

Nuclear power plants are important components of many energy systems, providing low-emission and economically competitive energy to communities [1]. With the operation and decommissioning of nuclear power plants, a large amount of nuclear waste is generated, and its effective disposal and management can be challenging. There are three main categories of radioactive waste: low-level waste (LLW), intermediate-level waste (ILW), and high-level waste (HLW). The classification system for radioactive waste can vary slightly depending on the region and specific regulatory frameworks. For example, the International Atomic Energy Agency (IAEA) classifies nuclear waste into five categories: Very-short-lived waste (VSLW), very-low-level waste (VLLW), LLW, ILW, and HLW [2]. The UK Nuclear Decommissioning Authority (NDA) uses the terms higher activity waste (HAW) and low activity waste (LAW). HAW includes HLW, ILW, and a small fraction of LLW that is not suitable for disposal to the LLW repository, while LAW includes LLW and VLLW [3]. The US Nuclear Regulatory Commission (NRC) regulates radioactive waste as LLW, HLW, Uranium Mill Tailings, and Waste incidental to reprocessing [4]. The Canadian Nuclear Safety Commission (CNSC) classifies all radioactive waste into four classes: HLW, ILW, LLW, and uranium mine and mill tailings. In the CNSC classification system, LLW contains two subclasses: VLLW and VSLW [5]. The summary of radioactive waste categories in different regions is shown in Figure 1. LLW refers to items that are contaminated by nuclear materials or exposed to neutron radiation and do not require shielding during handling and transportation [2,6]. It is typically characterised by a contact dose rate of less than 2 mSv/h, which is used to distinguish LLW from ILW and HLW [2]. LLW includes a diverse array of items with a wide range of radionuclides. Examples of LLW include routine cleaning supplies from nuclear power facilities, such as mopping heads, paper towels, and floor sweepings.
More than 90% of the total volume of radioactive waste is LLW or VLLW [7]. A significant proportion of LLW originates from the operation and decommissioning of nuclear facilities [3]. Nuclear decommissioning activities include actions that remove all or some regulatory control from nuclear facilities so that the site can be reused [8]. These activities will generate a large amount of predictable nuclear waste and can be challenging for local agencies. It is estimated that the radioactive waste inventory will increase globally in the next few decades due to predictable operational tasks in nuclear facilities, radiopharmaceuticals, commissioning activities, and new nuclear activities [3,7,9,10]. Handling the already substantial and continuously growing volume of LLW generated from the nuclear industry is challenging [11,12]. A typical LLW management process is shown in Figure 2.
Among various methods of handling LLW, sorting is the most effective approach to reducing waste volume [11,12]. Sorting involves the characterisation of LLW based on its radiological, chemical, and physical properties, and the classification of the waste into different streams (such as incinerable and metal) for further treatment processes. Sorting has proven to be a more sustainable approach to reducing the volume of LLW and extending the operational life of storage facilities [3]. After sorting LLW into different waste streams, there are three typical treatments (i.e., compaction, thermal treatment, and recycling) for waste volume reduction from the nuclear industry [13]. Compaction typically reduces the volume by 75% to 90%, and incineration can reduce the volume by more than 90% [3,14]. Recycling is another LLW treatment option to provide a second use for specific materials (e.g., metal), and it can achieve up to a 95% recycling rate for metallic waste [15]. While sorting presents a sustainable and promising solution to LLW management, it is still performed manually at most LLW handling facilities [16,17,18]. The automation of LLW sorting (especially for sorting based on physical properties) is in dire need to increase efficiency and cost savings, improve accuracy and scalability, and enhance operator safety from potential radiological exposure.
In the past decade, Computer Vision (CV) techniques demonstrated a significant potential for the automation of various sorting processes in different fields [19]. CV models use machine learning (ML) algorithms to segment and classify objects detected from images or videos. They can be integrated with other mechanical systems, such as robotic arms, to achieve automated sorting. CV has been applied in the solid waste management (SWM) domain with multiple applications, such as municipal waste detection and recognition [20]. The CV techniques and their applications in SWM have been summarised in previous studies [20,21,22]. However, LLW presents unique characteristics that make the direct adaptation of CV methods or CV-based SWM applications challenging. LLW has an extremely high standard for accuracy due to strict nuclear industry regulations and has less sorting destination categories compared to municipal solid waste. For these reasons, there is a need to summarise current progress and investigate the potential of CV applications for LLW sorting.
In this study, a comprehensive review of the previous work related to LLW sorting, including both scholarly and non-scholarly works, will be performed. A summative review of the existing CV applications in SWM and the state-of-the-art CV and ML algorithms for image recognition and classification will also be conducted. Finally, a discussion on the potential of CV and ML for automating LLW sorting will be presented, and a conceptual design for CV-aided LLW sorting systems and two potential designs for object classification components will be proposed. This work will provide a valuable reference for researchers, operators, and decision-makers, especially in LLW handling facilities, to facilitate efficient, sustainable, and innovative solutions for LLW handling.

2. Methodology

This review is designed to answer the following questions:
i) What are the existing LLW sorting technologies?
ii) What is the potential of CV techniques for LLW sorting?
A systematic review was first conducted to identify studies related to traditional LLW sorting and CV-aided LLW sorting (Figure 2). Several keywords, as well as their combinations, were used for a comprehensive search in the Web of Science (WoS) database. The number of publications was based on the results as of February 2025. All records, including their title, keywords, abstract, and full article, were reviewed manually. Additional searches were conducted on Scopus and Google Scholar to supplement the initial records. Seven publications on traditional LLW sorting and nine on nuclear waste management with CV techniques were selected for this review. Additionally, a non-scholarly search (Google) was also completed to review the latest industrial solutions for LLW sorting. The academic and non-scholarly search results are also summarised in Figure 3.
In addition, a summative review of CV and waste sorting automation was conducted to explore CV’s potential for LLW sorting. As there are many studies in the automatic sorting of non-radioactive materials, such as municipal solid waste, and numerous studies were conducted on the advancement of CV techniques, a summative review was performed to highlight the most relevant information from previous studies. This umbrella review also included records from both scholarly and non-scholarly searches. Two review papers focusing on SWM with CV techniques were selected. Additionally, one review paper on robotics in waste sorting and eleven records on the latest industry solutions for other solid waste sorting were included. Furthermore, three review papers on object detection and segmentation during video/image processing and three on multiclass classification were chosen to discuss the potential of CV for the automation of LLW sorting. The overview of the summative review is also shown in Figure 3.

3. Traditional LLW Sorting

Although a substantial number of articles on the policies and regulations for radioactive waste handling (approximately 25) were found through the literature search, there were very few that focused on the sorting processes, since the sorting processes are similar in different facilities [16,17,23,24]. Table 1 shows the sorting facilities found in the literature.
It is worth noting that these facilities are for general nuclear waste, not specifically for LLW. There is very limited literature focusing on LLW sorting, but the procedures are similar to those used for sorting other types of solid waste [16,23]. Nuclear waste materials are transported to processing facilities in containers (e.g., bins). At the processing facility, the containers are opened, and waste bags from the containers, usually one bag at a time, are removed from these containers and placed on a working station with protection (e.g., ventilation) [16,17]. The working station can be a sorting table or a glove box, depending on the facilities [26]. Waste materials from each bag can then be sorted into different categories (e.g., incinerable, metal) for further treatment or transportation to disposal by highly trained workers equipped with personal protective gear. In some cases, mechanical sorting systems were documented; however, they were specifically designed to handle particular waste streams, such as combustible and non-combustible materials, or to supplement manual sorting processes [24]. Although the sorting approaches in previous studies vary on a case-by-case basis, the process of LLW sorting reported in the literature is consistently described as manual and labour-intensive. The manual procedure requires extensive training for workers and a significant amount of time to sort each object individually, which not only limits efficiency but also introduces the possibility of human errors. Without an effective and efficient sorting solution, it can be challenging to handle the substantial volume of LLW and reduce storage demands, which is crucial for sustainable LLW management. Therefore, there is an urgent need for new, automated LLW sorting technology.

4. Recent CV Applications in Waste Management and Other Areas

4.1. Applications in Radioactive Waste Management

CV mimics the human visual system and performs tasks based on given vision data (e.g., images and videos) [19,27]. The latest CV techniques are often implemented with ML models to achieve the designed goals. In the literature, CV techniques were applied for radioactive waste handling only in a few studies, and they were used for waste sorting, object recognition, and scene reconstruction [28,29,30,31,32,33,34,35,36]. A list of the previous studies that utilised CV models is shown in Table 2.
Shaukat et al. [30] constructed an autonomous classification system based on the Suzuki contour algorithm and the random forest (RF) algorithm for nuclear decommissioning. Shaukat et al. [30] used a nuclear waste simulant dataset with a total of 86,400 objects from eight classes (i.e., bolt type A, bolt type B, head lifting eye, cone head, disc ring, roller follower, disc ring bolt, and nut type A). They applied feature detection processes (i.e., invariant feature detection and feature vector construction) and developed an RF-based classification system using Suzuki’s algorithm for contour extraction with an accuracy of approximately 98% [30]. Aitken et al. [28] designed an autonomous nuclear waste management system for decommissioning activities. This autonomous system included a KUKA KR500 robot arm (Augsburg, Germany) and used a time-of-flight camera to generate point clouds to detect and segment objects in waste canisters [28]. Kim et al. [29] developed a deep neural network (DNN) model for the characterisation of radioactive waste from the Korea Atomic Energy Research Institute. They extracted a total of 86,084 object images from categorisation videos and classified the objects into nine classes (i.e., vinyl, rubber, cotton, paper, plastic, wood, empty, hand-only, and unidentifiable) for model development. The developed model achieved an accuracy of 99.67% for single-object classification [29]. Sun et al. [31] proposed a weakly supervised approach for waste from nuclear decommissioning that can extract object information from RGBD videos, which capture both colour (RGB) and depth (D), with limited annotation (about 0.3% labelled data of the total dataset). The method showed better performance compared with other commonly used CV models, such as YOLO (You Only Look Once) v3 and Region-based convolutional neural network (CNN), in nuclear simulants recognition scenarios [31]. Zhao et al. [32] proposed a CV-aided system for material segmentation and reconstruction, which enables robots to work in complex environments and clean-up sites during nuclear decommissioning activities. This system used the Materials in Context (MINC) dataset, which included 23 categories of objects, for training, and used VGG-16 for object classification, fully convolutional networks (FCN), and conditional random fields as recurrent neural networks (CRF-RNN) for object segmentation, and graph-based simultaneous localisation and mapping (SLAM) algorithm for materials reconstruction [32]. Arhipov and Fomin [33] selected seven different advanced CNN-series models (i.e., VGG16, Inception V3, NASNetMobile, SqueezeNet, DenseNet121, MobileNetV2, MobileNet) for the classification of radioactive waste. They generated their dataset with six categories (i.e., cardboard, glass, metal, paper, plastic, textile, and wood) in a total of 15,038 images. Among the seven individual models evaluated, the Inception V3 achieved the highest accuracy at 87.56%; and the combination of the three best-performing models (i.e., Inception V3, DenseNet121, VGG 16) reached an accuracy of 91.58% [33]. Duani Rojas et al. [34] implemented five deep learning (DL) models for the classification of LLW. Based on their results, YOLO v7 was recommended due to its highest performance on detection and identification, and OWL-ViT was also recommended due to its zero-shot feature [34]. While not included in Table 2, two additional studies on scene reconstruction also demonstrated promising potential for enhancing LLW automation systems. Since sorting is influenced not only by visual features but also by radiological properties, these approaches may support integrated detection workflows in future design. Li et al. [35] proposed a point cloud fusion method combining binocular camera imagery and gamma radiation data to visualise surrounding environments. Tortajada et al. [36] presented a three-dimensional image reconstruction for gamma radiation detection for nuclear waste based on the portable geometry-independent tomographic system.
The data types of these radioactive waste handling systems can be divided into two categories: 2D and 3D. The former is constructed based on 2D object images and provides potential categories/classes of LLW during sorting. The latter is a constructed application based on 3D materials (e.g., RGB-D images). Meanwhile, there are two types of CV systems: integrated solutions and decision-support solutions. The difference between the two types of systems lies in the subsequent processes that follow the CV system. The downstream processes of the CV system in an integrated solution usually involve mechanical components for automatic sorting, such as robotic arms. This type of system usually requires 3D data for localisation. In the decision support solutions, the CV systems usually provide essential information (e.g., waste streams) of LLW objects to assist facility workers in manual sorting. Traditional ML (e.g., RF), CNN series (e.g., ResNet-50), and hybrid models (e.g., DCNN-GPC) have been applied in previous applications. For 2D decision-support applications involving fewer categories and objects, traditional ML and CNN models remain effective. For tasks with more complex conditions (e.g., 3D inputs and multi-model data), hybrid solutions are generally preferred.
A few more CV applications for nuclear waste handling were found through non-scholarly sources. For example, the UK NDA and UK Research and Innovation (UKRI) launched a competition called ‘Sort and Seg’ in 2021, aiming at developing novel techniques for handling nuclear waste generated from decommissioning activities [37]. All five winning proposals utilised robotics and ML techniques, and prototypes were made to implement their design (Table 3) [37,38,39].
In summary, limited efforts have been made to apply CV techniques to nuclear waste sorting, as evidenced by both academic and non-scholarly records. The achievements of demonstration from scholarly records prove the feasibility of using CV techniques for LLW automating sorting, while records from non-scholarly sources reflect that there is a need for automated LLW sorting on a commercial scale, and some companies are working on it. The aforementioned studies and prototypes demonstrate proof of concept for the innovative applications of CV in enhancing the efficiency of LLW sorting.

4.2. Applications for Other Solid Waste Management

Outside of the field of radioactive waste, CV has been widely utilised to sort other types of solid waste, and its applications have been summarised in previous studies [20,21]. An overview of three review papers relevant to the CV applications for solid waste handling is provided in Table 4.
Abdallah et al. [21] and Lu and Chen [20] systematically reviewed artificial intelligence (AI) and CV applications in SWM. Satav et al. [22] presented recent applications related to robotics in waste sorting and their implementations of waste detection and sorting mechanisms. Previous studies have summarised that CV and AI techniques are rapidly advancing in SWM and provide an effective alternative solution in this area [21]. Multiple types of AI or CV models (especially traditional ML and CNN-based models) were developed for waste detection, recognition, and classification with single and hybrid structures in SWM-related studies, and these models showed satisfactory performance in designed working conditions [20,21]. These models also have compatibility to integrate into the robotic systems for waste sorting tasks [22]. In addition, there are multiple solid waste object databases available, which makes it convenient for further exploration of CV techniques [20]. However, several common challenges still exist in this area, such as the lack of open and comprehensive datasets and oversimplified working conditions [20,21].
In industrial and commercial practices, a few automated solid waste sorting systems powered by AI techniques were developed, and eleven examples are selected and shown in Table 5.
These industrial applications can be classified into two categories: integrated solutions and decision-support solutions. The former includes physical sorting, i.e., mechanical modules, driven by CV-based object classification. Decision-support solutions, on the other hand, refer to standalone CV-based models or software that require further integration with either manual or mechanical sorting modules in the downstream of the waste handling process. There are two common types of mechanical sorting modules: conveyor belts and batch sorters. Conveyor belts can transport waste at a controlled speed, and they are typically designed for processing a large amount of waste continuously. Batch sorters involve manual or mechanical sorting and process waste on a batch basis, which can be a less efficient process compared to conveyor belts.
Although there are few existing designs specifically for the sorting of LLW, the aforementioned technologies used in automated solid waste systems provide valuable foundations for the development of automated sorting systems tailored to LLW. The difference between LLW and solid waste should be considered when adapting existing designs of solid waste automation systems to LLW sorting. For example, LLW has fewer types and destination streams compared to general solid waste, and it requires more accuracy. In addition, LLW is more sensitive than general solid waste as it requires more protocols to ensure radiological safety. Also, LLW sorting offers higher economic returns compared to general solid waste sorting, as the sorting and treatment of LLW can save space at dedicated and costly storage in disposal facilities. Moreover, the reduction in LLW is a major social benefit to the continued use and development of nuclear technology.

5. Potential of CV for Automatic Sorting

Currently, CV techniques are widely used in various areas, and previous researchers have summarised their applications of CV techniques in other areas. Table 6 lists a few literature review studies on state-of-the-art CV/ML techniques, which may have the potential for LLW sorting. CV has the potential to facilitate automatic sorting through two main tasks: object detection and segmentation in video/image processing and multiclass classification.
Object detection and segmentation in video/image processing are essential for creating customised LLW databases, which can facilitate the model training of CV-based LLW classification models and real-time LLW management. Because nuclear waste data is sensitive due to security, regulatory, and confidentiality concerns, operators and researchers often lack access to datasets from previous studies. As a result, they need to create LLW databases from scratch using videos and images from nuclear facilities to ensure accurate and secure data representation. Therefore, using CV techniques for object detection and segmentation on videos or images can significantly accelerate LLW database development. Sharma et al. [62] summarised common video processing tasks using DL techniques (e.g., object detection, event segmentation) based on publications from 2011 to 2020, and introduced widely used models for video processing and classified them into CNN-based, DNN-based, Recurrent Neural Network (RNN)-based, and hybrid approaches. Ilioudi et al. [63] presented existing DL techniques for the detection and segmentation of videos and outlined multiple directions for research in video processing (e.g., multimodal learning and representation learning with limited data). Le et al. [64] introduced several topics, such as object detection, image segmentation, and video analysis in CV using deep reinforcement learning. Previous studies demonstrated that CV models can detect and identify objects in videos and images. Given their capacity to recognise patterns between objects, it is reasonable to infer that LLW objects, which also exhibit distinct visual characteristics, can be similarly detected and extracted from videos and images by CV models. These publications [62,63,64] have stated the challenges of using DL and other ML techniques in CV, such as complicated environments and high computational resource requirements. Such issues are worth further investigation when CV models are applied to extract objects from LLW videos and images from nuclear facilities. Despite the challenges identified in previous studies, CV models for object detection and segmentation in video/image processing can automate the annotation process, which was traditionally a manual process. This presents significant potential for building an LLW database from scratch by employing the mentioned CV models in object detection and segmentation (e.g., CNN-based models, RNN-based models).
Meanwhile, multiclass classification based on CV is useful for classifying LLW objects into target streams. For LLW, there are several waste streams (e.g., incinerable, compactable, washable, metal, and non-processable) that aim to reduce the volume of LLW and can extend the operational life of disposal sites [13]. Ganesh et al. [65] summarised common DL-based CV models and introduced prominent areas with DL-based CV model applications (i.e., agriculture, health care, manufacturing, sports, and transportation), which included applications of object recognition for multiclass. Li et al. [67] analysed CV and DL applications on medical images in different areas of application, such as the brain and skin, and those applications included classification into different diseases. Geng et al. [66] summarised applications in three topics (i.e., parts recognition, agriculture picking, and intelligent driving) related to classification tasks and three primary challenges (i.e., small objects detection, stacked occlusion, real-time, and lightweight), and proposed solutions for each challenge. The aforementioned studies demonstrate that CV models can handle multiclass classification tasks in various areas (e.g., agriculture and manufacturing). Given the ability to classify between multiple categories with high accuracy, these models hold significant potential for adaptation to LLW classification scenarios.
DL models such as CNN and RNN are the most widely used methods for CV-related tasks, including object detection, segmentation, and recognition [62,63,65,66,67]. A typical CNN structure comprises multiple layers, including convolutional, pooling, and fully connected layers [20]. There are certain practices utilising CNN models (e.g., DCNN, VGG-16, ResNet-50) for nuclear waste classifications [29,31]. RNN is another powerful DL model for processing sequential data and has been mentioned multiple times in the collected literature [62,64,65,67]. A typical RNN structure comprises input layers, hidden layers, and output layers with a recurrent workflow [64]. Long short-term memory (LSTM) and gated recurrent units (GRU) are both representative models of RNN. Other DL approaches, including various unsupervised methods (e.g., autoencoders and Boltzmann machines), have the potential for LLW sorting [63].
Although Traditional ML models are not able to accept images as input directly, they can still perform computer vision classification tasks with the feature extraction process. Feature extraction is a process that transforms raw data (images) into numerical features for the original dataset, and it includes traditional image processing approaches (e.g., histogram of oriented gradients and scale-invariant feature transform) and DL (e.g., CNN) models [20]. Traditional ML models can process the data after the feature extraction. RF and SVM are widely used traditional ML models for feature classification tasks. RF is a supervised machine learning approach for classification and regression tasks. It uses a combination of decision trees to output the results based on the type of tasks (i.e., classification, regression), including the majority voting process or average value [68]. RF has potential as a model for LLW classification since it was used in one of the collected proof-of-concept publications for the classification of nuclear waste decommissioning tasks, and it was used in multiple SWM publications [20,21,30]. SVM is a supervised machine learning method, and the idea is to map an input vector to a high-dimensional feature space and construct a linear surface for decision-making [69]. SVM had been mentioned and described in multiple materials from our collections (i.e., SWM, other solid waste sorting, and medical images), mostly used as the classifier in a similar design [20,21,22,62,67]. Other traditional ML models, such as nearest neighbour, Bayesian network, linear regression, and artificial neural network, are also capable of performing LLW sorting tasks.

6. Conceptual Design for CV-Aided LLW Sorting Systems

Figure 4 illustrates a conceptual design for CV-aided LLW sorting systems for LLW handling facilities. It comprises three components: database development, object classification, and physical sorting. The database development process involves acquiring nuclear waste data from the facility, typically in the form of images and videos of LLW. The nuclear waste data are then processed to extract frames and object images, which are subsequently labelled with LLW stream type, matched with other relevant information (e.g., time, container ID), and then stored in an LLW database. The resulting images and annotation information can be managed in the LLW database to be used for CV model development and tuning.
The object classification process has two potential designs: direct classification and two-step classification. Direct classification uses an image-based classification model, which can directly output an object’s class (e.g., incineration, compaction, and metal) with the input of images or videos. This can be achieved through the development of DL models such as CNN and RNN. On the other hand, the two-step classification divides the object classification process into feature extraction and feature-based classification. As discussed in Section 5, the feature extraction process enables traditional ML models, which typically cannot process images or videos directly from the dataset, to facilitate the classification process. Additionally, it allows DL models to classify objects based on extracted features rather than directly using the images. Feature-based classification models include any models that can perform classification tasks based on extracted features. Traditional ML models and DL models are well-suited for such purposes.
In the third stage, after obtaining the objects’ position and class from the previous stages, the LLW object and its corresponding object information are processed downstream for physical sorting. As discussed in Section 4, physical sorting can be achieved through two approaches: integrated solutions and decision-support solutions. The integrated solution involves the use of automatic mechanical components to enable fully automated LLW sorting. The decision-support solution consolidates LLW object information and provides it to sorting technicians, assisting them in the sorting of LLW.

7. Challenges and Future Recommendations

Although there are existing studies on CV for automatic sorting and a few industrial applications of CV for waste handling, there are a few challenges to the future development of CV-aided automatic LLW sorting systems.
(i)
Data availability and quality: CV techniques, including ML and DL models, typically require large datasets for effective model training. The lack of high-quality datasets can result in poor model performance. Currently, there are very limited datasets of LLW objects, and almost all existing datasets in this area are confidential due to regulatory and security concerns within the nuclear sector. Guidelines and frameworks are needed to facilitate the creation of LLW databases and fill this critical data gap. These guidelines and frameworks should address the processing, formatting, storing, and managing of LLW data in various formats (e.g., object images, categorical and numeric waste information). Advanced LLW data processing tools that use CV techniques to detect and segment objects from LLW videos and images can also be useful.
(ii)
Model accuracy and generalisation capacity: Existing CV techniques commonly face challenges related to model accuracy. For example, overfitting can lead to high accuracy on the training set but poor performance on unseen LLW objects. Poor generalisation capability can result in poor performance in real-world scenarios with variations such as different lighting conditions, angles, or object occlusions. Meanwhile, LLW sorting typically involves multiple destination classes (e.g., incinerable, metal, compactable, non-processable, washable). Data imbalance can be a challenge to achieving satisfactory model accuracy for such multiclass classification tasks. Recently, more advanced CV techniques, such as data augmentation, fusion models, vision transformers, and generative adversarial networks, have shown potential for addressing these challenges in CV-aided LLW sorting and need to be further investigated.
(iii)
Lack of commercial applications: While there are a few pilot setups of CV-aided LLW sorting systems, they were developed in a laboratory environment and may not fully represent the complexities and scale of real-world applications. The use of CV techniques for LLW sorting needs to be further tested in real-world applications. Learning from the commercial applications of other automatic sorting systems outside of LLW, such as municipal solid waste and medical waste, can help bridge this gap and advance the development of commercial-scale systems.

8. Conclusions

This study presents a literature review on computer vision for low-level nuclear waste sorting. It provides a comprehensive review of both scholarly and non-scholarly works on traditional LLW sorting practices and the use of CV techniques in nuclear waste management. It also includes a summative review of the state-of-the-art techniques in waste sorting and CV. A conceptual design for CV-aided LLW sorting systems is presented, and the challenges and recommendations for the future development of CV-aided automatic LLW sorting systems are discussed.
Currently, the commercial LLW sorting process is mostly manual and labour-intensive. Recently, there have been several publications and prototypes focusing on nuclear waste handling. These solid waste automatic sorting applications provide the proof-of-concept for constructing LLW automatic sorting systems. Outside the field of waste management, CV techniques are widely used in numerous applications, which provide technical foundations for LLW automatic sorting. However, to the best of the authors’ knowledge, there are no large- or commercial-scale CV-aided applications for LLW sorting.
A conceptual design for CV-aided sorting systems with two potential designs of the object classification components, including direct classification and two-step classification, are presented. Direct classification utilises the image-based classification model that can accept images as input directly. DL models such as CNN and RNN are well-suited for this approach. On the other hand, the two-step classification system is based on feature-based classification models and needs to extract image features first. Then, algorithms capable of performing classification tasks based on the extracted features can be used to build the feature-based classification models. The LLW object information from the object classification process can be sent either to mechanical components or sorting technicians for waste sorting, depending on the type of application.
This study also identified future directions of CV-based LLW sorting applications. These included (a) developing high-quality LLW datasets for training models, (b) enhancing model accuracy and increasing generalisation capacity, and (c) constructing commercial applications for LLW sorting. Future research should focus on resolving those challenges.

Author Contributions

Conceptualisation, Z.L. and T.L.; methodology, T.L.; formal analysis, T.L.; investigation, T.L.; data curation, T.L.; writing—original draft preparation, T.L. and Z.L.; writing—review and editing, T.L., Z.L., and D.E.W.; visualisation, T.L.; supervision, Z.L. and D.E.W.; project administration, Z.L. and D.E.W.; funding acquisition, Z.L. All authors have read and agreed to the published version of the manuscript.

Funding

This research is funded by the Natural Sciences and Engineering Research Council of Canada (NSERC) and Laurentis Energy Partners (LEP).

Data Availability Statement

No data, models, or code were generated or used during the study.

Acknowledgments

The authors extend their sincere gratitude to their collaborators at LEP—Erik Rogerson, Alex Jay, Ryan Cooke, Daniel Orr, and Lionel Fernandes—for their invaluable contributions to this review paper.

Conflicts of Interest

The authors declare no conflicts of interest.

Abbreviations

The following abbreviations are used in this manuscript:
HLWHigh-level waste
ILWIntermediate-level waste
LLWLow-level waste
CVComputer vision
MLMachine learning
IAEAInternational Atomic Energy Agency
VSLWVery-short-lived waste
VLLWVery-low-level waste
NDANuclear Decommissioning Authority
HAWHigher activity waste
LAWLow activity waste
NRCNuclear Regulatory Commission
CNSCCanadian Nuclear Safety Commission
SWMSolid waste management
WoSWeb of Science
JETJoint European Torus
RFRandom forest
DNNDeep neural network
YOLOYou Only Look Once
CNNConvolutional neural network
MINCMaterials in Context
FCNFully convolutional networks
CRF-RNNConditional random fields as recurrent neural networks
SLAMSimultaneous localisation and mapping
DLDeep learning
UKRIUK Research and Innovation
BLSSSBarrnon Limited Sort and Segregate System
MASSMobile Autonomous Sort and Segregate System
AIArtificial intelligence
RNNRecurrent neural network
LSTMLong short-term memory
GRUGated recurrent units
NSERCNatural Sciences and Engineering Research Council of Canada
LEPLaurentis Energy Partners

References

  1. Jacobson, M. Review of solutions to global warming, air pollution, and energy security. Energy Environ. Sci. 2009, 2, 148–173. [Google Scholar] [CrossRef]
  2. IAEA. Classification of Radioactive Waste; International Atomic Energy Agency: Vienna, Austria, 2009; Available online: https://www.iaea.org/publications/8154/classification-of-radioactive-waste (accessed on 9 November 2023).
  3. UKRWI. 2022 Inventory Reports|UK Radioactive Waste Inventory (UKRWI). Available online: https://ukinventory.nda.gov.uk/the-2022-inventory/2022-inventory-reports/ (accessed on 4 February 2024).
  4. US NRC. Radioactive Waste. NRC Web. Available online: https://www.nrc.gov/waste.html (accessed on 22 April 2024).
  5. CNSC. Low- and Intermediate-Level Radioactive Waste. Available online: https://web.archive.org/web/20230721132156/https://nuclearsafety.gc.ca/eng/waste/low-and-intermediate-waste/index.cfm (accessed on 9 November 2023).
  6. US NRC. Low-Level Waste. NRC Web. Available online: https://www.nrc.gov/reading-rm/basic-ref/glossary/low-level-waste.html (accessed on 22 April 2024).
  7. IAEA. Status and Trends in Spent Fuel and Radioactive Waste Management; International Atomic Energy Agency: Vienna, Austria, 2022; Available online: https://www.iaea.org/publications/14739/status-and-trends-in-spent-fuel-and-radioactive-waste-management (accessed on 9 November 2023).
  8. IAEA. Global Status of Decommissioning of Nuclear Installations; International Atomic Energy Agency: Vienna, Austria, 2023; Available online: https://www.iaea.org/publications/15197/global-status-of-decommissioning-of-nuclear-installations (accessed on 25 February 2024).
  9. EC. Report from the Commission to the Council and the European Parliament on Progress of Implementation of Council Directive 2011/70/EURATOM and an Inventory of Radioactive Waste and Spent Fuel Present in the Community’s Territory and the Future Prospects Second Report. 2019. Available online: https://eur-lex.europa.eu/legal-content/EN/TXT/?qid=1581949250178&uri=CELEX:52019DC0632 (accessed on 4 February 2024).
  10. NRCan. Inventory of Radioactive Waste in Canada 2019; Natural Resources Canada: Ottawa, ON, Canada, 2021. [Google Scholar]
  11. ECCC. Reducing Municipal Solid Waste. Available online: https://www.canada.ca/en/environment-climate-change/services/managing-reducing-waste/municipal-solid/reducing.html (accessed on 14 August 2024).
  12. U.S. Environmental Protection Agency. Sustainable Materials Management: Non-Hazardous Materials and Waste Management Hierarchy. Available online: https://www.epa.gov/smm/sustainable-materials-management-non-hazardous-materials-and-waste-management-hierarchy (accessed on 14 August 2024).
  13. DECC. UK Strategy for the Management of Solid Low Level Waste from the Nuclear Industry. 2016. Available online: https://www.gov.uk/government/consultations/consultation-on-an-update-of-the-uk-strategy-for-the-management-of-solid-low-level-radioactive-waste-from-the-nuclear-industry (accessed on 14 May 2024).
  14. UKRWI. How Do We Manage Radioactive Waste?|UK Radioactive Waste Inventory (UKRWI). Available online: https://ukinventory.nda.gov.uk/about-radioactive-waste/how-do-we-manage-radioactive-waste/ (accessed on 5 January 2024).
  15. NDA. UK Strategy for the Management of Solid Low Level Radioactive Waste from the Nuclear Industry. 2010. Available online: https://assets.publishing.service.gov.uk/media/5a759463ed915d506ee7febb/UK_Strategy_for_the_Management_of_Solid_Low_Level_Radioactive_Waste_from_the_Nuclear_Industry_August_2010.pdf (accessed on 2 October 2024).
  16. Beer, H.-F. Radioactive waste management at the Paul Scherrer Institute, the largest Swiss national research centre. Nucl. Technol. Radiat. Prot. 2009, 24, 212–218. [Google Scholar] [CrossRef]
  17. Casey, S.M. Inspection of Low-Level Radioactive Waste. Hum. Factors 1991, 33, 1–15. [Google Scholar] [CrossRef]
  18. Reynolds, S.; Newman, M.; Coombs, D.; Witts, D. JET experience on managing radioactive waste and implications for ITER. Fusion Eng. Des. 2016, 109–111, 979–985. [Google Scholar] [CrossRef]
  19. Shapiro, L.G.; Stockman, G.C. Computer Vision; Prentice Hall: Upper Saddle River, NJ, USA, 2001. [Google Scholar]
  20. Lu, W.; Chen, J. Computer vision for solid waste sorting: A critical review of academic research. Waste Manag. 2022, 142, 29–43. [Google Scholar] [CrossRef] [PubMed]
  21. Abdallah, M.; Abu Talib, M.; Feroz, S.; Nasir, Q.; Abdalla, H.; Mahfood, B. Artificial intelligence applications in solid waste management: A systematic research review. Waste Manag. 2020, 109, 231–246. [Google Scholar] [CrossRef]
  22. Satav, A.G.; Kubade, S.; Amrutkar, C.; Arya, G.; Pawar, A. A state-of-the-art review on robotics in waste sorting: Scope and challenges. Int. J. Interact. Des. Manuf. 2023, 17, 2789–2806. [Google Scholar] [CrossRef]
  23. Clark, D.E.; Lerch, R.E. Volume Reduction Options for the Management of Low-Level Radioactive Wastes; HEDL-SA-1235; Hanford Engineering Development Lab: Richland, WA, USA, 1977. Available online: https://www.osti.gov/biblio/7296686 (accessed on 20 November 2023).
  24. IAEA. Treatment of Low and Intermediate Level Solid Radioactive Wastes; International Atomic Energy Agency: Vienna, Austria, 1983; Available online: https://www.iaea.org/publications/1324/treatment-of-low-and-intermediate-level-solid-radioactive-wastes (accessed on 15 August 2024).
  25. Shropshire, D.; Sherick, M.; Biadgi, C. Waste Management Facilities Cost Information for Low-Level Waste; INEL-95/0013; Lockheed Idaho Technologies Co.: Idaho Falls, ID, USA, 1995. [Google Scholar] [CrossRef]
  26. Newbert, G.; Haigh, A.; Atkins, G. The Handling, Assessment, Transport and Disposal of Tritiated Waste Materials at JET. Fusion Technol. 1995, 28, 1552–1557. [Google Scholar] [CrossRef]
  27. Huang, T.S. Computer Vision: Evolution and Promise; CERN European Organization for Nuclear: Geneva, Switzerland, 1996. [Google Scholar]
  28. Aitken, J.M.; Veres, S.M.; Shaukat, A.; Gao, Y.; Cucco, E.; Dennis, L.A.; Fisher, M.; Kuo, J.A.; Robinson, T.; Mort, P.E. Autonomous Nuclear Waste Management. IEEE Intell. Syst. 2018, 33, 47–55. [Google Scholar] [CrossRef]
  29. Kim, J.-G.; Jang, S.-C.; Kang, I.-S.; Lee, D.-J.; Lee, J.-W.; Park, H.-S. A study on object recognition using deep learning for optimizing categorization of radioactive waste. Prog. Nucl. Energy 2020, 130, 103528. [Google Scholar] [CrossRef]
  30. Shaukat, A.; Gao, Y.; Kuo, J.A.; Bowen, B.A.; Mort, P.E. Visual classification of waste material for nuclear decommissioning. Robot. Auton. Syst. 2016, 75, 365–378. [Google Scholar] [CrossRef]
  31. Sun, L.; Zhao, C.; Yan, Z.; Liu, P.; Duckett, T.; Stolkin, R. A Novel Weakly-Supervised Approach for RGB-D-Based Nuclear Waste Object Detection. IEEE Sens. J. 2019, 19, 3487–3500. [Google Scholar] [CrossRef]
  32. Zhao, C.; Sun, L.; Stolkin, R. Simultaneous Material Segmentation and 3D Reconstruction in Industrial Scenarios. Front. Robot. AI 2020, 7, 52. [Google Scholar] [CrossRef]
  33. Arhipov, A.; Fomin, I. Research of Convolutional Neural Networks in the Problem of Classification of Images of Radioactive Waste. In Proceedings of the 2022 International Conference on Industrial Engineering, Applications and Manufacturing (ICIEAM), Sochi, Russia, 16–20 May 2022; pp. 854–858. [Google Scholar] [CrossRef]
  34. Duani Rojas, A.; Lagos, L.; Upadhyay, H.; Soni, J.; Prabakar, N. AI-based detection and identification of low-level nuclear waste: A comparative analysis. Neural Comput. Applic 2024, 36, 21061–21072. [Google Scholar] [CrossRef]
  35. Li, Y.; Cheng, Z.; Yang, C.; Wei, M.; Wen, J. Application of Binocular Stereo Vision in Radioactive Source Image Reconstruction and Multimodal Imaging Fusion. IEEE Trans. Nucl. Sci. 2020, 67, 2454–2462. [Google Scholar] [CrossRef]
  36. Tortajada, S.; Albiol, F.; Caballero, L.; Albiol, A.; Leganés-Nieto, J.L. A portable geometry-independent tomographic system for gamma-ray, a next generation of nuclear waste characterization. Sci. Rep. 2023, 13, 12284. [Google Scholar] [CrossRef] [PubMed]
  37. NDA. Five UK Companies Have Progressed Through to the Next Phase of a £5.5 Million Radioactive Waste Segregation Competition. GOV.UK. Available online: https://www.gov.uk/government/news/five-uk-companies-have-progressed-through-to-the-next-phase-of-a-5-5-million-radioactive-waste-segregation-competition (accessed on 10 November 2023).
  38. NDA. Results of Competition: Sort and Segregate Nuclear Waste: Phase 1. 2021. Available online: https://assets.publishing.service.gov.uk/media/607d9cbee90e076aa7da98ca/Competition_Results_-_Sort_and_Segregate_Nuclear_Waste_-_Phase_1.pdf (accessed on 10 November 2023).
  39. NDA. Sort and Segregate Nuclear Waste Competition. 2023. Available online: https://www.youtube.com/watch?v=tsa9hx2cL_U (accessed on 15 February 2024).
  40. UKRI. Barrnon Limited Integrated Sort and Segregate Solution-All Databases. 2021. Available online: https://www.webofscience.com/wos/alldb/full-record/GRANTS:15380240 (accessed on 11 February 2024).
  41. Barrnon Nuclear. BLISSS—Artificial Intelligence for Nuclear Decommissioning. 2023. Available online: https://www.youtube.com/watch?v=8WxhgA1wIvw (accessed on 25 February 2024).
  42. UKRI. ISOsort-Grants Index. 2021. Available online: https://www.webofscience.com/wos/grants/full-record/GRANTS:15400581 (accessed on 11 February 2024).
  43. Createc. ISOSort—Sorting and Segregating Nuclear Waste. 2023. Available online: https://www.youtube.com/watch?v=-oZWN6PFJuA (accessed on 11 February 2024).
  44. UKRI. Mobile Autonomous Sort and Segregate System-Grants Index. 2021. Available online: https://www.webofscience.com/wos/grants/full-record/GRANTS:15336354 (accessed on 11 February 2024).
  45. ATKINS. Mobile, Autonomous Sort and Segregation. 2022. Available online: https://www.gamechangers.technology/static/u/Atkins.pdf (accessed on 11 February 2024).
  46. UKRI. OptiSort-Grants Index. 2021. Available online: https://www.webofscience.com/wos/grants/full-record/GRANTS:15312494 (accessed on 11 February 2024).
  47. UKRI. Blended Intelligence for Safe and Efficient Nuclear Sort & Segmentation-Grants Index. Available online: https://www.webofscience.com/wos/grants/full-record/GRANTS:15392190 (accessed on 11 February 2024).
  48. Veolia Nuclear Solutions. The Blended Intelligence for Safe and Efficient Nuclear Sort and Segregation. 2023. Available online: https://www.youtube.com/watch?v=GQa0bwcDu9Y (accessed on 11 February 2024).
  49. MSWSorting. Sorting Robot for Waste Recycling—MSWsorting. Available online: https://www.mswsorting.com/Waste-Sorting/Sorting-Robot.html (accessed on 15 March 2024).
  50. MSWsorting. Construction and Demolition Waste Sorting Robot, Building Waste AI Sorting Robot, www.MSWsorting.com. 2022. Available online: https://www.youtube.com/watch?v=RUtiPf1CEW8 (accessed on 21 March 2024).
  51. Dadlani, J. IRS Waste Sorting Robotics. 2019. Available online: https://www.youtube.com/watch?v=6EQwA2CIVio (accessed on 14 March 2024).
  52. World of Bin-e. Bin-e—The Smartest Waste Bin Worldwide. 2020. Available online: https://www.youtube.com/watch?v=LJYWRTRJThY (accessed on 15 March 2024).
  53. CleanRobotics. TrashBot: The Smart Recycling Bin That Sorts at the Point of Disposal, CleanRobotics. Available online: https://web.archive.org/web/20240229211742/https://cleanrobotics.com/trashbot/ (accessed on 15 March 2024).
  54. Bulk Handling Systems (BHS). Max AI® Technology Overview. 2022. Available online: https://www.youtube.com/watch?v=B99RpOOlfP8 (accessed on 14 March 2024).
  55. Bulk Handling Systems (BHS). Empowering Advanced & Mechanical Recycling: Max-AI Contribution to Sustainable Polyolefin Sorting. Max-AI®. Available online: https://max-ai.com/empowering-advanced-mechanical-recycling-max-ai-contribution-to-sustainable-polyolefin-sorting/ (accessed on 14 March 2024).
  56. AMP. Home. Available online: https://ampsortation.com/ (accessed on 15 March 2024).
  57. Greyparrot, AI. Automate Waste Analytics with Greyparrot Analyser|Greyparrot AI. Available online: https://www.greyparrot.ai/products/analyser (accessed on 15 March 2024).
  58. TOMRA. GAINTM—Deep Learning. Available online: https://www.tomra.com/en/waste-metal-recycling/products/technologies/gain (accessed on 15 March 2024).
  59. ZenRobotics. ZenBrain, ZenRobotics. Available online: https://www.terex.com/zenrobotics/robots/zenbrain (accessed on 15 March 2024).
  60. Recycleye. Home, Recycleye. Available online: https://recycleye.com/ (accessed on 15 March 2024).
  61. WasteAnt. Home, WasteAnt. Available online: https://wasteant.com/ (accessed on 14 March 2024).
  62. Sharma, V.; Gupta, M.; Kumar, A.; Mishra, D. Video Processing Using Deep Learning Techniques: A Systematic Literature Review. IEEE Access 2021, 9, 139489–139507. [Google Scholar] [CrossRef]
  63. Ilioudi, A.; Dabiri, A.; Wolf, B.J.; De Schutter, B. Deep Learning for Object Detection and Segmentation in Videos: Toward an Integration With Domain Knowledge. IEEE Access 2022, 10, 34562–34576. [Google Scholar] [CrossRef]
  64. Le, N.; Rathour, V.S.; Yamazaki, K.; Luu, K.; Savvides, M. Deep reinforcement learning in computer vision: A comprehensive survey. Artif. Intell. Rev. 2022, 55, 2733–2819. [Google Scholar] [CrossRef]
  65. Ganesh, N.; Shankar, R.; Mahdal, M.; Murugan, J.; Chohan, J.; Kalita, K. Exploring Deep Learning Methods for Computer Vision Applications across Multiple Sectors: Challenges and Future Trends. Comput. Model. Eng. Sci. 2023, 139, 103–141. [Google Scholar] [CrossRef]
  66. Geng, K.; Qiao, J.; Liu, N.; Yang, Z.; Zhang, R.; Li, H. Research on Real-time Detection of Stacked Objects Based on Deep Learning. J. Intell. Robot. Syst. 2023, 109, 82. [Google Scholar] [CrossRef]
  67. Li, J.; Jiang, P.; An, Q.; Wang, G.-G.; Kong, H.-F. Medical image identification methods: A review. Comput. Biol. Med. 2024, 169, 107777. [Google Scholar] [CrossRef] [PubMed]
  68. Breiman, L. Random Forests. Mach. Learn. 2001, 45, 5–32. [Google Scholar] [CrossRef]
  69. Cortes, C.; Vapnik, V. Support-vector networks. Mach. Learn. 1995, 20, 273–297. [Google Scholar] [CrossRef]
Figure 1. Radioactive waste categories in different areas.
Figure 1. Radioactive waste categories in different areas.
Environments 12 00270 g001
Figure 2. A typical LLW management process.
Figure 2. A typical LLW management process.
Environments 12 00270 g002
Figure 3. Overview of the academic and non-scholarly search results (n denotes the number of records).
Figure 3. Overview of the academic and non-scholarly search results (n denotes the number of records).
Environments 12 00270 g003
Figure 4. A conceptual design for CV-aided LLW sorting systems.
Figure 4. A conceptual design for CV-aided LLW sorting systems.
Environments 12 00270 g004
Table 1. Traditional nuclear waste sorting facilities.
Table 1. Traditional nuclear waste sorting facilities.
ReferenceFacility/InstituteRegion/CountrySorting Equipment
Clark and Lerch [23]Unknown/ClassifiedNorth AmericaGlove box or walk-in room with pressurised suit
IAEA [24]Unknown/ClassifiedUnknown/ClassifiedGlove box
IAEA [24]Unknown/ClassifiedUKPressurised suit areas
IAEA [24]Unknown/ClassifiedNetherlandsVibration sieve
IAEA [24]Unknown/ClassifiedUnknown/ClassifiedAir classifier
Casey [17]Unknown/ClassifiedUSASorting tables with hood, viewing window, and ventilation system
Shropshire et al. [25]Unknown/ClassifiedUSASorting tables
Newbert et al. [26]Joint European Torus (JET)UKBoth sorting tables and glove boxes with ventilation system
Beer [16]Paul Scherrer InstituteSwitzerlandWalk-in cells with manipulator
Reynolds et al. [18]JETUKEnclosed waste processing boxes with ventilation system
Table 2. Previous CV systems for radioactive waste handling.
Table 2. Previous CV systems for radioactive waste handling.
ReferenceNameData TypeSolution TypeAlgorithmApplication ScenarioAccuracyDatasetNote
Shaukat et al. [30]A vision-based autonomous sorting-and-segregation system2DDecision-supportSuzuki contour algorithm, RFRedundant nuclear waste material classification from nuclear decommissioning process98.15%Nuclear waste simulants dataset from National Nuclear Laboratory and Sellafield Ltd., Warrington, UK.
  • Proved CV’s capability to assist in nuclear decommissioning waste sorting.
Aitken et al. [28]A nuclear waste autonomous robotic treatment system 3DIntegrated Random sample consensus, Euclidean cluster extractionRemote processing and export of redundant nuclear waste prior to downstream encapsulation//
  • Processed randomly and continuously presented undefined waste from canisters.
  • Provided a video demonstration of sorting process.
Sun et al. [31]A weakly supervised learning model for nuclear decommissioning waste3DIntegratedDCNN-GPC Waste objects detection and categorisation for nuclear decommissioning Washington RGB-D object recognition dataset: 86.2% (RGB), 76.3% (Depth), 90.2% (RGB-D); Birmingham nuclear waste simulants dataset: 80.9% (instant-wise detection), 75.5% (pixel-wise detection)Birmingham nuclear waste simulants dataset, Washington RGB-D object recognition dataset
  • Achieved better performance compared to other baseline models (i.e., YOLO v3, 3D R-CNN).
Zhao et al. [32]A real-time material segmentation and 3D reconstruction system3DIntegratedVGG-16, FCN, CRF-RNN, graph-based SLAMRecognition of nuclear waste material types MINC dataset: 81.94% (pixel accuracy), 74.19% (mean accuracy); Industrial scenario: 80.10% (pixel accuracy), 58.75% (mean accuracy),MINC dataset
  • Conducted in a real-world industry room containing multiple materials.
Kim et al. [29]A DL-based radioactive recognition system 2DDecision-supportResNet-50Categorisation of nuclear waste99.67%Nuclear waste dataset from the sorting process
  • Established a dataset of manually categorised nuclear wastes.
Arhipov and Fomin [33]CNN models in radioactive waste classification2DDecision-supportVGG16, Inception V3, NASNetMobile, SqueezeNet, DenseNet121, MobileNetV2, MobileNetCategorisation of nuclear waste91.58% (combined model), 87.56% (Inception V3)Custom waste dataset
  • Inception V3 performed the best compared to other CNN-series models.
  • The best-performed combination models contained Inception V3, DenseNet121, VGG 16.
Duani Rojas et al. [34]DL models on LLW detection and identification2DDecision-supportYOLO v7, STEGO, SD Mask-RCNN, OWL-ViTLLW detection and identification0.995 (mAP, YOLOv7-seg) Custom LLW dataset
  • Generated a dataset based on 38 images.
  • Pre-trained with COCO and WISDOM datasets.
Table 3. Summary of the five prototypes in the UK NDA sort and seg competition.
Table 3. Summary of the five prototypes in the UK NDA sort and seg competition.
CompanyProduct Name 1Key FeaturesReferences
Barrnon Ltd. (Westmorland, UK)Barrnon Limited Sort and Segregate System (BLSSS)A fully robotic system that identifies, sorts, and segregates nuclear waste using ML techniques.[40,41]
Createc Ltd. (Cockermouth, UK)ISOsortAn automatic robot picking, segregation, sorting, and characterisation system for nuclear waste.[42,43]
Atkins Ltd. (Epsom, UK)Mobile Autonomous Sort and Segregate System (MASS)A fully autonomous robotic system that sorts and segregates nuclear waste via CV.[44,45]
Cavendish Nuclear Ltd. (Warrington, UK)OptiSortAn autonomous waste sorting, segregation, and packing system for nuclear waste[46]
Veolia Nuclear Solutions Ltd. (Abingdon, UK)Blended Intelligence for Safe and Efficient Nuclear Sort and SegmentationA touch-sensitive remote manipulator platform for nuclear waste sorting and segmentation[47,48]
1 The “product name” column refers to the hardware and software system mentioned in each cited record.
Table 4. Previous review studies on CV applications in solid waste management.
Table 4. Previous review studies on CV applications in solid waste management.
ReferenceReview FocusNumber of Papers Included 1CV/ML ModelsScope
Abdallah et al. [21]General SWM and AI techniques84Traditional ML models, CNN-based models, and hybrid models 2004–2019
Lu and Chen [20]Computer Vision with SWM87Traditional ML models, CNN-based models, and hybrid models1997–2021
Satav et al. [22]Robotics in Waste Sorting32Traditional ML models, CNN-based models, and CV algorithms2010–2023
1 The “number of papers included” column refers to the number of studies analysed in each cited review paper.
Table 5. Examples of automated solid waste sorting systems.
Table 5. Examples of automated solid waste sorting systems.
CompanyProduct NameKey FeaturesSorting ScenarioSolution TypeReference
MSW technology (Zhengzhou, China)MSW SortingGrabs recyclable waste from mixed garbage based on DL algorithmsConveyor beltIntegrated solution[49,50]
Ishitva Robotic Systems (Gujarat, India)YUTAHas robots that sort waste; uses decision-making analytics based on CV techniquesConveyor beltIntegrated solution[51]
Bine sp. z o. o. (Dąbrowa, Poland)Bin-eSorts and compresses waste based on AI techniquesBatch sorterIntegrated solution[52]
CleanRobotics (Longmont, CO, USA)TrashBotClassifies waste based on ML algorithmsBatch sorterIntegrated solution[53]
Bulk Handling Systems (Eugene, OR, USA)Max AIIdentifies waste; integrated with multiple sorting componentsConveyor beltIntegrated solution[54,55]
AMP (Louisville, CO, USA)AMP ONESorts waste by AI techniquesConveyor beltIntegrated solution[56]
Greyparrot AI (London, UK)Greyparrot SyncIntegrates and enhances sorting components based on AI techniquesConveyor beltSoftware package[57]
TOMRA (Mülheim-Kärlich, Germany)GAINnextSorts waste based on DL techniquesConveyor beltIntegrated solution[58]
ZenRobotics (Helsinki, Finland)ZenBrainIs trained to sort 500 categories of waste, with increased waste sorting efficiency Conveyor beltIntegrated solution, software package[59]
Recycleye (London, UK)Recycleye InsightsMonitors and analyses all items on the belt based on AI techniquesConveyor beltIntegrated solution, software package[60]
WasteAnt (Bremen, Germany)Conveyor beltMonitors all items on the belt based on AI techniquesConveyor beltSoftware package[61]
Table 6. Previous review studies on the application of CV and ML models.
Table 6. Previous review studies on the application of CV and ML models.
ReferenceReview FocusNumber of Papers IncludedCV/ML ModelsScope
Sharma et al. [62]Video processing in:
Human action recognition
Motion detection
Object detection/recognition/tracking
Video classification
Behaviour analysis
Gait analysis
Background subtraction
Event recognition
Action segmentation
Scene understanding
93CNN-based
DNN-based
RNN-based
Hybrid approach
2011–2020
Ilioudi et al. [63]Video processing in:
Semantic segmentation
Classification and localisation
Object detection
Instance segmentation
/CNN
Restricted Boltzmann machine
Autoencoder
LSTM
GRU
Self-attention mechanism
/
Le et al. [64]Landmark localisation
Object detection/tracking
Image registration
Image segmentation
Video analysis
/DL
Reinforcement learning
/
Ganesh et al. [65]DL and CV in
Agriculture
Healthcare
Manufacturing
Sports
Transportation
/CNN
RNN
Deep belief networks
Deep Boltzmann machine
Deep energy models
Autoencoders
/
Geng et al. [66]Real-time detection of stacked objects/Backbone network
Region proposal network
2015–2022
Li et al. [67]Computer vision and medical image18CNN
DL
ML
Transfer learning
2014–2022
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Li, T.; Winckler, D.E.; Li, Z. Computer Vision for Low-Level Nuclear Waste Sorting: A Review. Environments 2025, 12, 270. https://doi.org/10.3390/environments12080270

AMA Style

Li T, Winckler DE, Li Z. Computer Vision for Low-Level Nuclear Waste Sorting: A Review. Environments. 2025; 12(8):270. https://doi.org/10.3390/environments12080270

Chicago/Turabian Style

Li, Tianshuo, Danielle E. Winckler, and Zhong Li. 2025. "Computer Vision for Low-Level Nuclear Waste Sorting: A Review" Environments 12, no. 8: 270. https://doi.org/10.3390/environments12080270

APA Style

Li, T., Winckler, D. E., & Li, Z. (2025). Computer Vision for Low-Level Nuclear Waste Sorting: A Review. Environments, 12(8), 270. https://doi.org/10.3390/environments12080270

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop