A Lightweight Detection Method for Meretrix Based on an Improved YOLOv8 Algorithm

Abstract

Clam farms are typically located in remote areas with limited computational resources, making it challenging to deploy traditional deep learning-based object detection methods due to their large model size and high computational demands. To address this issue, this paper proposes a lightweight detection method, YOLOv8-RFD, based on an improved YOLOv8 algorithm, tailored for clam sorting applications. The proposed enhancements include the following: replacing the original backbone network of YOLOv8 with a Reversible Columnar Network (RevColNet) to reduce feature redundancy and computational load; upgrading the C2f modules in both the backbone and neck networks to C2f-Faster to optimize feature fusion strategies and improve fusion efficiency; and incorporating a Dynamic Head (DyHead) to enhance feature extraction and detection accuracy by adaptively adjusting the detection head structure. Experimental results on a custom clam dataset demonstrate that, compared to the original YOLOv8 model, the proposed method reduces the number of parameters by 22.75% and computational demand by 18.52%, while slightly improving detection accuracy. These improvements not only maintain but also enhance detection performance, significantly reducing computational cost, and confirming the method’s suitability for deployment in resource-constrained environments. This provides a reliable technical foundation for the sorting of clams.

1. Introduction

The clam is widely cultivated due to its rapid growth and high economic value; however, its complex and variable morphological characteristics pose challenges for efficient identification and classification in large-scale farming and automated processing [1,2]. Clam detection models are commonly deployed on edge devices such as sorting machines and mobile terminals. They are widely used in scenarios like aquaculture farms and seafood processing lines, placing higher demands on detection accuracy and deployment performance. Therefore, it is of great significance to develop a target detection model that balances both high detection accuracy and lightweight design.

In recent years, artificial intelligence technologies, particularly deep learning and convolutional neural networks (CNNs), have achieved remarkable progress in image recognition and object detection, driving their widespread adoption in the aquaculture industry. To improve detection accuracy, Tu Xueying et al. [3] addressed the challenges associated with the small size, high transparency, and irregular morphology of Paralichthys olivaceus larvae, achieving precise counting through the application of the ResNet34 model. Ji et al. [4] enhanced feature extraction by employing an improved MobileNetv2 backbone integrated with a coordinate attention module, resulting in a 3.2% increase in mean average precision (mAP) on a crab dataset. Wang et al. [5] developed an improved YOLOv3 architecture (YOLO-D) by reinforcing the residual module, incorporating a dual attention mechanism, and constructing a secondary recursive feature pyramid network, thereby enhancing the detection accuracy of small objects in offshore UAV imagery. Similarly, Zhang Hongjiao et al. [6] optimized the DenseNet framework by integrating the SENet attention mechanism and refining the network structure, achieving a recognition accuracy of 91.53% on a custom dataset. Addressing the challenge of small target omission and false detection in aerial imagery, Li et al. [7] integrated Bi-PAN-FPN for multi-scale feature fusion into YOLOv8-s, utilized GhostblockV2 to reduce parameters and suppress information loss, and optimized anchor boxes through the WiseIoU method and a dynamic non-monotonic focusing mechanism, resulting in notable performance improvements on public datasets. Regarding model lightweighting, Hao et al. [8] proposed an efficient object detection algorithm for external damage on power transmission lines, leveraging ACmix and ShuffleNetv2 to substantially reduce model parameters and computational load, while integrating the PC-ELAN module and SIOU loss function for enhanced performance. Mao et al. [9] developed a lightweight SAR ship detector by combining a simplified U-Net architecture with an anchor-free detection framework, achieving model compression at the expense of some detection accuracy. Additionally, Ma et al. [10] presented the YOLOv8-HD model, which employed optimized detection heads and deformable attention mechanisms to enhance feature extraction, achieving 77.6% mAP in stacked wheat seed scenarios while reducing model size by 20%. Sun et al. [11] developed the lightweight YOLOv5-PRE model for apple detection, combining ShuffleNet and GhostNet to minimize model size and enhancing feature extraction with CA and CBAM attention mechanisms, ultimately achieving 94.03% accuracy with a 17.93% increase in detection speed. Xin et al. [12] proposed YOLOv7-SDBB, a lightweight underwater object detection method that employs the ShuffleNetv2 backbone to reduce computational load, integrates D-ELAN and D-MPConv modules to accelerate detection, and utilizes BiFPN and BiFormer for improved feature fusion and accuracy. Validation on public datasets confirmed substantial reductions in model parameters and computation, while maintaining high accuracy, thereby achieving an effective balance between precision and efficiency.

Currently, research on the detection and recognition of clam remains limited. The external features of clams are subtle and easily affected by environmental factors such as silt and water, and the presence of morphologically similar shellfish in the detection environment further complicates accurate identification. Consequently, models with high detection precision typically require substantial computational resources, which constrains their practical deployment. To address these challenges, achieving model lightweighting while maintaining high performance becomes particularly critical. In this study, we propose a lightweight detection method for clams based on an improved YOLOv8 framework, aiming to enhance both detection accuracy and deployment efficiency.

2. Materials and Methods

2.1. Dataset Preparation

The images of clams were captured using a 12-megapixel camera. The data collection scenarios cover four typical distribution patterns: single target with dispersed placement, single target with clustered placement, multiple targets with dispersed placement, and multiple targets with clustered placement (Figure 1). This setup is designed to simulate the actual working environment of sorting machine systems, focusing primarily on challenges such as occlusion caused by stacked targets, interference from impure shellfish, and object detection under varying densities. A total of 225 images were collected for each scenario, resulting in an initial dataset of 900 images. All images were saved in the “jpg” format, and each batch of samples was randomly arranged to ensure data diversity and representativeness.

To enhance the model’s performance and generalization ability, data augmentation techniques were employed to expand the sample set and mitigate the risk of overfitting caused by insufficient training data. In this study, random pixel dropout, image sharpening, affine transformation, brightness adjustment, hue adjustment, and horizontal flipping were randomly combined as data augmentation methods [13] to enrich the dataset. Using these techniques, five new augmented images were generated from each original image, with some examples shown in Figure 2. As a result, a total of 3000 image samples were produced through data augmentation. The dataset was then randomly divided into training, validation, and test sets in a ratio of 8:1:1.

2.2. Experimental Environment and Parameter Settings

The experiments in this study were conducted based on the following hardware and software configurations: the hardware platform employed a high-performance computing system equipped with an Intel Xeon E5-2689 v4 processor (3.10 GHz clock speed, manufactured by Intel Corporation, Santa Clara, CA, USA), an NVIDIA GeForce RTX 3090 GPU (24 GB memory, manufactured by NVIDIA Corporation, Santa Clara, CA, USA), and 512 GB of RAM. The software environment was built on the Windows 10 operating system, utilizing the Python 3.8 programming language, CUDA 11.6 parallel computing architecture, and the PyTorch 1.13.1 deep learning framework. For model training, RGB images with a resolution of 640 × 640 pixels were used as input. The stochastic gradient descent (SGD) optimizer was applied with an initial learning rate of 0.01, a momentum factor of 0.937, and weight decay (coefficient of 0.0005) to prevent overfitting. The experiments employed mini-batch gradient descent with a batch size of 8 and a total of 50 epochs, balancing computational efficiency with model convergence stability.

2.3. Improved YOLOv8 Network

The YOLO (You Only Look Once) network is a popular real-time object detection system first proposed by Joseph Redmon and colleagues in 2016 [14]. The design of this network allows it to detect objects in an image with a single scan, contrasting sharply with previous methods that required multiple scans. The YOLOv8 series models offer five versions—n, s, m, l, and x—tailored for different application scenarios, with increasing model depth and detection accuracy. Among them, YOLOv8n features the smallest model size and the highest detection speed, effectively meeting the demands for real-time performance and lightweight architecture. Given the focus on model deployment efficiency in this study, YOLOv8n was selected as the base framework for optimization.

To address the challenges of high model complexity and insufficient recognition accuracy in current clam detection technologies, this study proposes a lightweight optimization detection solution based on YOLOv8. The improved model introduces enhancements to the backbone, neck, and detection head of YOLOv8. The architecture of the improved model is shown in Figure 3.

Firstly, since the current YOLO series models adopt a top-down structure in their backbone networks, this design tends to lose embedded information during the feature extraction process, which can lead to a decline in model performance. To address this issue, this paper proposes a reconstructed backbone network based on a Reversible Connected Multi-Column Network (RevColNet) [15]. RevColNet breaks through the traditional information transmission mode of feedforward networks by adopting a multi-input design, where each column starts with low-level information and extracts semantic information through compressed image channels. Reversible connections are used between columns to ensure lossless data transmission and to endow the network with reversibility. Supervision is added at the end of each column to constrain feature extraction. To prevent excessive complexity in the backbone network and avoid increasing model complexity and parameter count, the number of columns in RevCol is set to 2. Additionally, the operations within the Fusion Block are restructured. For high-level semantic information, downsampling is performed using convolution, batch normalization, and activation functions; for low-level semantic information, convolution and upsampling are applied. Furthermore, the ConvNeXt module in YOLOv8 is replaced with the C2f module.

Secondly, although the C2f module used in YOLOv8 enhances the image feature extraction capability, the stacking of Bottleneck modules inevitably leads to redundant information channels and increased inference workload. To address these issues, the Faster Block module is integrated into C2f, thereby reducing the model’s computational cost and floating-point operations (FLOPs). This structure adopts the Partial Convolution (PConv) mechanism, where convolution is applied only to a subset of channels while the remaining channels remain unchanged, effectively reducing redundant computations and memory access overhead. This approach significantly lowers computational complexity and FLOPs while maintaining—or even enhancing—the model’s feature representation capability and detection performance, demonstrating superior inference efficiency and deployment advantages.

Finally, to improve the model’s adaptability to small objects and complex backgrounds, the DyHead structure is adopted to replace the original detection head in YOLOv8. DyHead dynamically fuses multi-scale features, enabling the network to adaptively handle objects of varying sizes. This significantly enhances detection performance for multi-scale targets such as clams and improves robustness to variations in object scale. Moreover, DyHead optimizes the feature integration strategy, enriching semantic information and improving localization accuracy while reducing false positives in background regions. As a result, it demonstrates strong adaptability and high accuracy when dealing with complex scenarios and diverse detection challenges.

3. Improved Lightweight Method for Clam Detection

3.1. RevColNet

RevColNet (Reversible Connected Multi-Column Network) is a lightweight backbone architecture designed to enhance feature representation while reducing information loss. It achieves lossless cross-layer feature transmission through a multi-column design and reversible connection mechanism, improving the fusion of contextual information. This makes it particularly suitable for small object detection tasks, where preserving low-level details is crucial [16]. The macro structure of RevColNet is shown in Figure 4.

The microarchitecture of RevColNet is depicted in Figure 5. As shown in Figure 5a, each layer performs feature extraction through a combination of downsampling and ConvNeXt blocks. Figure 5b illustrates the design of reversible connections between columns, which enable bidirectional information exchange and preserve feature integrity throughout the network. Each layer receives two inputs: one from the preceding layer within the same column, and another from the corresponding layer of the adjacent column. The formal representations of these inputs are provided in Equations (1) and (2).

$$X_t=F_t\left(X_{t-1},X_{t-m+1}\right)+\gamma X_{t-m}$$

(1)

$$X_{t-m}={\gamma }^{(-1)}[X_t-F_t(X_{t-1},X_{t-m+1})]$$

(2)

In the equations, $X_t$ represents the feature map at layer t, $F_t(.)$ denotes the activation function, $\gamma$ refers to the reversible operation, and ${\gamma }^{-1}$ is its inverse. Each column consists of m feature maps within the group.

3.2. C2f-Faster

C2f-Faster is a lightweight improved version of the C2f module in YOLOv8. By introducing the Partial Convolution (PConv) mechanism, it performs convolution operations on only a subset of feature channels, effectively reducing FLOPs and memory access. This significantly lowers the model’s computational load while maintaining feature extraction performance [17]. PConv reduces redundant computations and memory accesses, achieving higher processing speeds without compromising accuracy across various vision tasks. Assuming that the number of output channels remains the same and k is the kernel size, the formulas for the floating-point operations (FLOPs) and memory access cost (MAC) of PConv are as follows [18,19]:

$$FLOP{s}_{(PConv)}=h\times w\times k^2\times c_p^2$$

(3)

$$MAC=h\times w\times 2c_p+k^2\times c_p^2\approx h\times w\times 2c_p$$

(4)

The module performs convolution operations on a subset of the input channels c_p, representing the entire feature map, while leaving the remaining channels unchanged. The processed channels are then concatenated with the unprocessed ones and passed as the output. Under typical Partial Convolution rates (r = 1/4), the computational cost of the improved C2f-Faster model is approximately 1/16 of the original C2f, with significantly lower memory usage during the convolution process. Compared to conventional convolution, the memory access is about 1/4 of that required by the standard convolution operation. The structural layout of the C2f-Faster module is illustrated in Figure 6.

Therefore, this study adopts the FasterBlock from FasterNet to replace the Bottleneck module in the C2f of YOLOv8, achieving a lightweight network architecture.

3.3. DyHead

DyHead (Dynamic Head) is a multi-dimensional dynamic attention mechanism detection head that integrates scale-aware, spatial-aware, and task-aware mechanisms, enabling adaptive fusion of multi-scale features [20]. Figure 7 shows the structure of the DyHead module. The detection head of the YOLOv8n model is improved to enable scale-aware dynamic detection head attention, with the corresponding formula for the attention function calculation as follows:

$$W\left(F\right)={\pi }_C({\pi }_S({\pi }_L(F)\cdot F)\cdot F)\cdot F$$

(5)

In the formula, ${\pi }_C$ represents task-aware attention, ${\pi }_S$ represents spatial-aware attention, ${\pi }_L$ represents scale-aware attention, and F represents the feature vector.

By introducing the self-attention mechanism in three specific dimensions, the detection head’s ability to perceive clam features is enhanced, improving the algorithm’s detection performance on small targets. This includes scale-aware attention, spatial-aware attention, and task-aware attention, and their respective calculation formulas are as follows:

$${\pi }_L(F)\cdot F=\sigma (f({\displaystyle \frac{1}{SC}}{\displaystyle {\sum }_{SC}F}))F$$

(6)

$${\pi }_s(F)\cdot F={\displaystyle \frac{1}{L}}{\displaystyle {\sum }_{l=1}^L{\displaystyle {\sum }_{k=1}^K{w}_{l,k}\cdot F(l;p_k+\Delta p_k;c)\cdot \Delta }}{m}_k$$

(7)

$${\pi }_C(F)\cdot F=max({\alpha }^1(F)\cdot F_c+{\beta }^1(F),{\alpha }^2(F)\cdot F_c)+{\beta }^2(F)$$

(8)

In the formulas, $\sigma$ represents the activation function, $f$ represents the transformation function, $P_k$ represents the position, $\Delta P_k$ represents the position offset, $\Delta m_k$ represents the additional weight, $K$ represents the number of key points, and $F_c$ represents the features segmented on the C-th channel.

The scale-aware attention module enhances multi-scale feature representation ability by dynamically fusing semantic features of different scales. This module introduces a linear function f(·), equivalent to a 1 × 1 convolution operation, and combines it with the hard-sigmoid activation function σ(x) = max(0, min(1, x + 1/2)) to efficiently filter features. The spatial-aware attention module further strengthens the consistency modeling of spatial position information and feature hierarchy on top of feature fusion. It uses deformable convolution to sparsify the attention distribution, focusing on the cross-layer feature relationships at the same spatial location. The task-aware attention module dynamically adjusts the activation state of feature channels based on specific task requirements to enhance the model’s adaptability to multi-task scenarios.

4. Experiment and Analysis

4.1. Evaluation Metrics

In this study, the performance of the model is evaluated using the mAP50, mAP50:95, number of parameters, and computational power (GFLOPs). mAP50 represents the average detection accuracy for each category when the Intersection over Union (IoU) threshold is set to 0.5. mAP50:95 measures the average detection accuracy across all IoU thresholds ranging from 0.5 to 0.95 (with a step size of 0.05), providing a more comprehensive reflection of the model’s performance. The calculation formula for mAP is as follows:

$$Precision={\displaystyle \frac{TP}{TP+FP}}$$

(9)

$$Recall={\displaystyle \frac{TP}{TP+FN}}$$

(10)

$$AP={\displaystyle {\int }_0^{1}p(r)dr}$$

(11)

$$mAP={\displaystyle \frac{1}{n}}{\displaystyle {\sum }_{i=1}^{n}A{P}_i}$$

(12)

In the formulas, TP (True Positive) refers to the number of detection boxes where IoU > 0.5, indicating the number of clams correctly identified by the model; FP (False Positive) refers to the number of detection boxes where IoU ≤ 0.5, indicating the number of clams incorrectly identified by the model; and FN (False Negative) refers to the number of clams not identified by the model. Using the above evaluation metrics, the algorithm’s performance can be comprehensively assessed in terms of detection accuracy, computational efficiency, and model lightweighting, providing a basis for subsequent optimization.

4.2. Experimental Results and Analysis

4.2.1. Comparison of Different Algorithms

To verify the effectiveness of the improved YOLOv8 algorithm proposed in this study, a series of evaluation metrics were adopted, including precision, recall, mAP@0.5, mAP@0.5:0.95, the number of parameters, and GFLOPs. A comparative experiment was conducted with Faster R-CNN, YOLOv5s, and YOLOv8n on the self-constructed clam dataset. All algorithms were tested under the same experimental conditions, with the batch size set to 8 and the total number of epochs set to 50. The Faster R-CNN algorithm used ResNet50 as its backbone network. The experimental results are presented in

Table 1. Comparison of experimental results.

Model	Precision/%	mAP0.5/%	mAP0.5:0.95/%	Params/M	GFLOPs
Faster R-CNN	99.5	98.1	74.6	41.4	239.3
YOLOv5s	85.4	84.5	62.1	7.1	16.5
YOLOv8n	89.8	89.3	70.4	3.2	8.9
YOLOv8-RFD	89.9	89.5	72.2	2.2	6.6

.

Based on the experimental results presented in Table 1, it can be concluded that although the two-stage object detection algorithm Faster R-CNN demonstrates outstanding detection accuracy, it suffers from high computational complexity and a large number of parameters, resulting in excessively large model weight files. Therefore, it is not suitable for the lightweight and real-time detection requirements of the dataset used in this study. In contrast, YOLOv8n outperforms YOLOv5s in terms of parameter count and computational efficiency, while also achieving slightly higher detection accuracy, indicating better overall performance. The proposed YOLOv8-RFD algorithm achieves comparable detection accuracy to YOLOv8n, but shows improved performance under the more stringent mAP@0.5:0.95 metric, while significantly reducing both the number of parameters and computational cost. In summary, the improved YOLOv8 algorithm proposed in this study demonstrates superiority across multiple key metrics and is better suited for lightweight real-time detection tasks.

4.2.2. Ablation Study

To verify the effectiveness of the proposed improvement strategies, this study conducted ablation experiments using YOLOv8n as the baseline algorithm on a self-constructed clam dataset.

Table 2. Results of ablation experiment.

Model	Parameter	GFLOPs	mAP50/%	mAP50:95/%
YOLOv8n	3,157,200	8.9	89.3	70.4
YOLOv8n + DyHead	3,485,263	9.6	89.6	74.0
YOLOv8n + RevCol	2,276,435	6.3	89.2	70.6
YOLOv8n + C2f-Faster	2,300,643	6.3	89.2	70.4
YOLOv8n + RevCol + DyHead	2,755,855	7.8	89.3	71.7
YOLOv8n + RevCol + C2f-Faster	1,751,835	5.0	89.1	67.2
YOLOv8n + C2f-Faster + DyHead	2,780,063	7.9	89.5	72.4
YOLOv8n + RevCol + C2f-Faster + DyHead	2,231,255	6.6	89.5	72.2

presents the comparative results of each improved module, and the variation curve of mAP@0.5:0.95 is illustrated in Figure 8.

The experimental results show that the introduction of the Dynamic Head (DyHead) led to an increase of 0.3% in mAP@0.50 and 3.6% in mAP@0.50:0.95. However, the number of parameters increased from 3,157,200 to 3,485,263, and the GFLOPs rose by 0.7. This indicates that while DyHead significantly enhances the model’s accuracy in clam detection, it also markedly increases model complexity. To balance model performance and computational cost, this study replaced the backbone network with RevColNet and adopted the more efficient C2f-Faster processing strategy, thereby significantly reducing model complexity while ensuring improvements in both mAP@0.50 and mAP@0.50:0.95. When RevColNet and C2f-Faster were applied simultaneously, mAP@0.50 reached 89.5%, and mAP@0.50:0.95 improved to 72.2%, demonstrating a compounded performance gain from the combined use of these components. These results strongly validate the effectiveness and feasibility of the proposed strategies in enhancing clam detection performance while maintaining a low computational cost.

4.2.3. Visualization Experiments

To assess the detection performance of the improved model, representative data from dispersed placement, clustered placement, and multi-object scenarios were selected for visualization experiments. The comparison results between the improved YOLOv8 algorithm and the original YOLOv8 are shown in Figure 9. The results indicate that the improved YOLOv8 algorithm demonstrates significant advantages in both recognition accuracy and bounding box regression performance. Particularly in multi-object stacking scenarios, the original YOLOv8 algorithm tends to make erroneous detections and exhibits lower recognition accuracy, while the improved model effectively mitigates issues related to low accuracy and misidentification of clams. Additionally, the improved algorithm can accurately distinguish similar targets in multi-object scenarios, significantly improving the precision of clam target localization and feature recognition. These experimental results further demonstrate the practicality of the improved algorithm in complex scenarios.

5. Conclusions

This paper proposes a lightweight YOLOv8 algorithm (YOLOv8-RFD) based on a deep reversible architecture, which improves detection accuracy while reducing resource consumption. The aim is to efficiently detect and recognize clams in resource-constrained environments. The experimental results validate the applicability of the algorithm in clam detection. The main conclusions are as follows:

(1)

In the feature extraction stage, RevColNet is introduced as the backbone network, replacing the original deep convolutional structure. This network enhances the representation ability for clam targets through its efficient feature aggregation mechanism. It exhibits stronger feature learning ability when handling complex background interference and multi-scale targets, while significantly reducing redundant computation. The introduction of RevCol alone reduces the number of parameters from 3.16 M to 2.28 M, while maintaining the mAP@50:95 without any degradation. It is suitable for target detection tasks in complex background conditions.

(2)

In terms of model lightweighting, the C2f-Faster lightweight convolution module is adopted to replace the traditional feature fusion unit. This module optimizes the computation path and parameter reuse strategy, reducing the GFLOPs to 6.3 with minimal loss in accuracy.

(3)

In the design of the detection head, the DyHead dynamic detection head is introduced. Through the dynamic allocation mechanism of cross-layer feature weights, the model’s adaptability to multi-scale targets is enhanced. Compared with the traditional fixed weight allocation method, this approach captures target edge information more accurately, further reducing false detection rates and enabling small target detection in dense scenes.