Estimating Recycling Return of Integrated Circuits Using Computer Vision on Printed Circuit Boards

: The technological growth of the last decades has brought many improvements in daily life, but also concerns on how to deal with electronic waste. Electrical and electronic equipment waste is the fastest-growing rate in the industrialized world. One of the elements of electronic equipment is the printed circuit board (PCB) and almost every electronic equipment has a PCB inside it. While waste PCB (WPCB) recycling may result in the recovery of potentially precious materials and the reuse of some components, it is a challenging task because its composition diversity requires a cautious pre-processing stage to achieve optimal recycling outcomes. Our research focused on proposing a method to evaluate the economic feasibility of recycling integrated circuits (ICs) from WPCB. The proposed method can help decide whether to dismantle a separate WPCB before the physical or mechanical recycling process and consists of estimating the IC area from a WPCB, calculating the IC’s weight using surface density, and estimating how much metal can be recovered by recycling those ICs. To estimate the IC area in a WPCB, we used a state-of-the-art object detection deep learning model (YOLO) and the PCB DSLR image dataset to detect the WPCB’s ICs. Regarding IC detection, the best result was obtained with the partitioned analysis of each image through a sliding window, thus creating new images of smaller dimensions, reaching 86.77% mAP. As a ﬁnal result, we estimate that the Deep PCB Dataset has a total of 1079.18 g of ICs, from which it would be possible to recover at least 909.94 g of metals and silicon elements from all WPCBs’ ICs. Since there is a high variability in the compositions of WPCBs, it is possible to calculate the gross income for each WPCB and use it as a decision criterion for the type of pre-processing.


Introduction
Waste electrical and electronic equipment (WEEE) management is a concern in many countries [1] since electrical and electronic equipment production has increased while the average lifetime of these products has decreased [2]. Consequently, WEEE has the fastest growth rate in the industrialized world [3]. This concern has led to the development of legislation and directives to manage this sort of waste, for example, the European Directive for WEEE and the Brazilian Policy of Solid Waste (BPSW) [4].
The European Commission recently adopted Implementing Regulation (EU) 2019/290, which focuses on restricting the use of hazardous substances in electrical and electronic equipment in the WEEE Directive of 2003. The evolution of laws and regulations demonstrates an advancement regarding the impacts that the improper disposal of electronic materials can have on the environment, thus the development of projects in this line of research positively reflects the industry's current sustainability issues.
Printed circuit boards (PCB) are an essential element present in almost all electrical and electronic equipment [2]. The PCB is the platform that connects the discrete and 2 of 11 integrated electronic components. Recycling waste printed circuit boards (WPCBs) can recover valuable materials (metallic and nonmetallic), besides the possibility of the reuse of some components [5]. WPCBs are the major source of metals in electronic waste [6]. Even given the high-value materials, WPCB recycling is difficult because of the diversity in their compositions [7,8].
Dismantling and separating of the WPCB components is very important to achieve the maximum recycling outcomes. However, completely automating this pre-processing is a high-cost process, and it is performed manually in primitive methods in some countries [9]. In order to optimize the recycling WPCB process, vision systems can be used to input data to desoldering systems and identify reusable and hazardous components [7,10]. Aiming to promote computer-vision-based strategies for PCB recycling, the PCB DSLR public dataset was published [10]. DSLR stands for Digital Single-Lens Reflex camera, and it is a reference of the camera type used to acquire the PCB images. The PCB DSLR Dataset contains high-resolution WPCB images with bounding box annotations for integrated circuits (ICs). Since some ICs may contain valuable materials, it is essential to detect such components in WPCBs.
Computer vision systems are a feasible way to perform WPCB evaluation to guide automatic dismantling and recycling [11]. The computer vision task performed is named object detection. Since sufficiently large fully-annotated datasets for the PCB domain are expensive and unavailable [12], the use of state-of-the-art object detectors (based on deep learning) can be impaired because of the dataset size [13]. This sort of problem is known as few-shot learning [14]. The main goal is to learn proper feature embedding that allows the model to be trained with fewer images than standard models.
Since ICs are a significant source of highly valuable metals, we propose the WPCB economic feasibility assessment (WPCB-EFA) method for IC recycling from a WPCB by estimating recoverable metals and gross earnings using computer vision to detect ICs in WPCBs. As WPCBs have a high composition diversity, they can be grouped according to their value. For instance, mobile phones and mainframe PCBs have a high value (8 to 25 euros per kg) while calculators, audio scraps, and power supplies are low-value PCBs (less than 1 euro per kg) [15]. The proposed WPCB-EFA can be used to evaluate the composition of a WPCB and select the appropriated recycling process focused on recovering precious metals for high-value PCBs or the nonmetal fraction for low-value PCBs.
To evaluate the proposed WPCB-EFA, we used transfer learning of the state-of-the-art YOLOv3 model [16] in a few-shot scenario for IC detection on the PCB DSLR Dataset. Then, we proposed an estimation for the IC surface density to calculate the PCB's IC weight and estimate the financial return of recycling a WPCB's integrated circuits.
The rest of this article is structured as follows. Section 2 provides an overall context about WPCB, IC recycling, and object detection. In Section 3, there is a detailed description of the proposed method for the economic evaluation of IC recycling from WPCBs. Section 4 presents the results of the proposed method applied to the PCB DSLR Dataset. Finally, Section 5 contains the final discussion and our conclusions.

Printed Circuit Board (PCB) and Integrated Circuit (IC) Recycling
The PCB composition is complex, so the recycling process is arduous [7,9]. A PCB comprises about 30% metallic components. Some of these are high-value metals like gold, silver, and copper. Even though these high-value materials are less than 1% of the PBC weight, it can represent 80% to 88% of the PCB recycling value [9,15]. Li et al. [7] estimated that 1 ton of WPCBs contained approximately 284 g of gold.
WPCB recycling processes are mainly interested in the recovery of those valuable metals. However, depending on the recycling process used, there are environmental problems with residues and hazardous substances resulting from the recycling process [17,18].
Usually, the WPCB recycling process involves three steps: dismantling, physical crushing or mechanical process, and some chemical leaching processes [9]. The dismantling Table 1. Physical processing of the discarded integrated circuit (IC) outcomes [20]. Aside from physical separation, leaching is another approach for integrated circuit recycling [21]. The leaching process is applied after physical separation. Lee at. al. reported 100% recovery of gold (Au), silver (Ag), and copper (Cu) with thiourea leaching [21].

Recovered Elements for 1 kg of ICs
Considering recycling the entire WPCB, precious metals (Au, Pd, Pt, and Ag) are about 88% of the intrinsic value of the WPCB [15]. Most studies consider the recovery of these precious metals from printed circuit boards as a whole [15]. Although studies indicate the concentration of the precious metal in integrated circuits, the percentage of the integrated circuit mass corresponding to precious metals is challenging [21]. This measurement difficulty occurs because the concentration of precious metals is small, and there is a considerable variation in concentration depending on the sample of the ICs.

Object Detection
Object detection is a computer vision task to determine an object's location by outputting a bounding box and, at the same time, classifying the object [22]. When multiple objects are detected on a particular image, this is referred to as multi-class object detection.
Deep learning strategies are state-of-the-art for object detection problems [22]. These strategies can be split into two main classes: two-stage and one-stage methods. The twostage methods, the R-CNN (region with convolutional neural network features) family, first generate region proposals and then classify each region. The one-stage method (YOLO model) executes the object detection task in a single forward through the neural network.
YOLO stands for You Only Look Once, and it is a real-time object detection system. The general idea is that a single neural network divides the image into regions and then predicts several bounding boxes for each region. Its probability selects those bounding boxes as a final result [16].
This work used the third version of YOLO (YOLOv3), which has improved performance compared to earlier versions. Since the first version, YOLO's backbone for feature extraction and classification is a deep convolutional network. In YOLOv2, the backbone was Darknet-19, while in YOLOv3, it was replaced by Darknet-53, a deeper network. Both Darknet architectures are convolutional neural networks. YOLOv3 also replaced the final softmax layer used in YOLOv2 for independent logistic classifiers (with binary cross-entropy loss). This change made the model become a multilabel object detector. Figure 1 shows the outline of the proposed WPCB economic feasibility assessment (WPCB-EFA). The first step is to detect the ICs and then calculate the IC area and weight. With the IC weight, it is possible to use the proposed IC recycling methods to estimate the recoverable metals and the gross earnings [20,21].

Waste Printed Circuit Boards Economic Feasibility Assessment (WPCB-EFA)
The general idea is that a single neural network divides the image into predicts several bounding boxes for each region. Its probability select boxes as a final result [16].
This work used the third version of YOLO (YOLOv3), which has mance compared to earlier versions. Since the first version, YOLO's bac extraction and classification is a deep convolutional network. In YOLO was Darknet-19, while in YOLOv3, it was replaced by Darknet-53, a dee Darknet architectures are convolutional neural networks. YOLOv3 also softmax layer used in YOLOv2 for independent logistic classifiers (with tropy loss). This change made the model become a multilabel object det Figure 1 shows the outline of the proposed WPCB economic feas (WPCB-EFA). The first step is to detect the ICs and then calculate the IC With the IC weight, it is possible to use the proposed IC recycling metho recoverable metals and the gross earnings [20,21].

Detect IC Regions
To detect the ICs in the WPCB images, we evaluated the transfer from the pre-trained YOLOv3 model. In order to deal with the few-sh lems, one approach is to use prior knowledge and perform a data aug dure [14]. Therefore, we conducted data augmentation in the PCB DSLR ate smaller sub-images. A sliding window of 416 × 416 pixels generat passing through the original image. The training process used only sub-i

Detect IC Regions
To detect the ICs in the WPCB images, we evaluated the transfer learning strategy from the pre-trained YOLOv3 model. In order to deal with the few-shot learning problems, one approach is to use prior knowledge and perform a data augmentation procedure [14]. Therefore, we conducted data augmentation in the PCB DSLR dataset to generate smaller sub-images. A sliding window of 416 × 416 pixels generated sub-images by passing through the original image. The training process used only sub-images containing at least a fraction of some IC. After data augmentation, the new dataset had 5347 RGB images containing ICs, each one with a size of 416 × 416 pixels.
Since the IC detector works with chunks of full-resolution WPCB images, the reverse process is necessary to obtain the entire WPCB image's IC bounding boxes. Given that the WPCB will be in a conveyor belt, we defined the region of interest as the conveyor belt area. Figure 2 illustrates this region of interest. In summary, we removed the region outside the conveyor belt. Since all of the PCB DSLR images were from the same perspective, the region of interest's coordinates were defined to remove the region outside the conveyor belt.
area. Figure 2 illustrates this region of interest. In summary, w side the conveyor belt. Since all of the PCB DSLR images were f the region of interest's coordinates were defined to remove th veyor belt. Therefore, we applied a sliding window approach in the r to the process of building the training set. The YOLO model pro gathered by the sliding window. Then, the YOLO prediction for bined over the original size image to obtain the bounding bo image. Figure 3 shows the prediction result for a full-size imag taset. The sliding window stride allows multiple bounding box were concerned with the IC area, standard non-maximum supp did not apply to this problem. Therefore, all the predicted bou higher than the threshold (0.8) were used to produce an image IC image masks for the YOLO predictions and ground truth b the IC image mask, the IC pixel area was calculated. Therefore, we applied a sliding window approach in the region of interest, similarly to the process of building the training set. The YOLO model processes every image chunk gathered by the sliding window. Then, the YOLO prediction for each image chunk is combined over the original size image to obtain the bounding boxes over the entire WPCB image. Figure 3 shows the prediction result for a full-size image from the PCB DSLR Dataset.
area. Figure 2 illustrates this region of interest. In summary, we removed the reg side the conveyor belt. Since all of the PCB DSLR images were from the same pers the region of interest's coordinates were defined to remove the region outside veyor belt. Therefore, we applied a sliding window approach in the region of interest, s to the process of building the training set. The YOLO model processes every imag gathered by the sliding window. Then, the YOLO prediction for each image chunk bined over the original size image to obtain the bounding boxes over the entir image. Figure 3 shows the prediction result for a full-size image from the PCB D taset. The sliding window stride allows multiple bounding boxes for the same IC were concerned with the IC area, standard non-maximum suppression (NMS) alg did not apply to this problem. Therefore, all the predicted bounding boxes with higher than the threshold (0.8) were used to produce an image mask. Figure 3 sh IC image masks for the YOLO predictions and ground truth bounding boxes. B the IC image mask, the IC pixel area was calculated.

Calculating IC Weight
With a WPCB image, it is possible to estimate the surface area of each in circuit. However, to calculate the IC's weight mass, we needed to estimate the surface density. Texas Instruments (TI) has published an Application Report wi mation on the IC weight and footprint area of fifteen IC packages [23]. Figure 4 sh calculated surface density for each IC package and the average IC surface densi The sliding window stride allows multiple bounding boxes for the same IC. As we were concerned with the IC area, standard non-maximum suppression (NMS) algorithms did not apply to this problem. Therefore, all the predicted bounding boxes with a score higher than the threshold (0.8) were used to produce an image mask. Figure 3 shows the IC image masks for the YOLO predictions and ground truth bounding boxes. Based on the IC image mask, the IC pixel area was calculated.

Calculating IC Weight
With a WPCB image, it is possible to estimate the surface area of each integrated circuit. However, to calculate the IC's weight mass, we needed to estimate the average surface density. Texas Instruments (TI) has published an Application Report with information on the IC weight and footprint area of fifteen IC packages [23]. Figure 4 shows the calculated surface density for each IC package and the average IC surface density based on the TI application report. Since most of the IC weight can be assigned to the package, we calculated the average surface density for an IC of 2358 mg/mm 2 .
Appl. Sci. 2021, 11, x FOR PEER REVIEW 6 of 11 on the TI application report. Since most of the IC weight can be assigned to the package, we calculated the average surface density for an IC of 2358 mg/mm 2 .

Recycling Economic Evaluation
The general economic evaluation of WPCB recycling was proposed by Niu et al. [2]. The total income for metal powder is given by , where is the net income per kilogram metal powder; is the commodity cost of the metal element; is the composition of that particular metal; denotes the total weight number; and is the recycling cost. The recycling cost is highly dependent on the specific production line, so it must be explicitly calculated for one industrial plant. A complete discussion regarding the recycling cost is beyond this paper's scope since it involves analyzing static and dynamic costs. Therefore, we can use Equation (1) to calculate the gross profit of IC recycling. Since we are conducting economic evaluation based on visual analysis, the IC weight needs to be estimated from the IC area. So, the IC weight is given by (2) where is the IC weight; is the IC surface area in pixels; is the scale factor from pixel to mm ; and is the IC surface density in g/mm With the IC weight, the gross recycling profit can be calculated as , where is the composition of metal (obtained from Table 1); is the total ICs weight; and is the metal market value.

Experimental Setup
For the development of the IC object detector, we used the PCB DSLR Dataset [10]. Initially, this dataset has 748 RBG images of WPCBs, with a resolution of 4928 × 3280 pixels, from 165 different boards (three to five images per board). Furthermore, the integrated circuits present on the boards are all labeled, totaling 9313 components.
The WPCBs were extracted from a recycling facility, and some WPCBs had broken parts and dust all over them. The WPCBs had different sizes and shapes, as can be seen in Figure 5. This makes the dataset very close to what is encountered in a real-world scenario.

Recycling Economic Evaluation
The general economic evaluation of WPCB recycling was proposed by Niu et al. [2]. The total income for metal powder is given by where G is the net income per kilogram metal powder; C i is the commodity cost of the i th metal element; P i is the composition of that particular metal; η denotes the total weight number; and M is the recycling cost. The recycling cost is highly dependent on the specific production line, so it must be explicitly calculated for one industrial plant. A complete discussion regarding the recycling cost is beyond this paper's scope since it involves analyzing static and dynamic costs. Therefore, we can use Equation (1) to calculate the gross profit of IC recycling. Since we are conducting economic evaluation based on visual analysis, the IC weight needs to be estimated from the IC area. So, the IC weight is given by where W is the IC weight; A is the IC surface area in pixels; S is the scale factor from pixel to mm 2 ; and ρ is the IC surface density in g/mm 2 With the IC weight, the gross recycling profit can be calculated as where P i is the composition of i th metal (obtained from Table 1); W is the total ICs weight; and C i is the metal market value.

Experimental Setup
For the development of the IC object detector, we used the PCB DSLR Dataset [10]. Initially, this dataset has 748 RBG images of WPCBs, with a resolution of 4928 × 3280 pixels, from 165 different boards (three to five images per board). Furthermore, the integrated circuits present on the boards are all labeled, totaling 9313 components.
The WPCBs were extracted from a recycling facility, and some WPCBs had broken parts and dust all over them. The WPCBs had different sizes and shapes, as can be seen in Figure 5. This makes the dataset very close to what is encountered in a real-world scenario. For each PCB, the dataset has images in different positions. This is important because usually, WPCBs are placed on conveyor belts with arbitrary orientation, and a computer-vision system has to perform visual inspection despite the PCB orientation.
sion system has to perform visual inspection despite the PCB orientation.
In order to obtain the IC surface area, we needed to evaluate the dataset image scale. The silkscreen printing easily identified some IC part numbers on the package. With the measurement of the IC dimensions present in the datasheet, it is possible to calculate the real object size scale. It is important to note that the dataset authors do not provide information about the camera calibration. Table 2 shows the pixel area estimate for a particular IC (part number RTL810L).  For all experiments, we used the pre-trained YOLOv3 provided by Darknet [16]. The fine-tune strategy was to restore the weights from all layers except the last three convolutional layers. Additionally, the YOLOv3 head was fine-tuned with a momentum optimizer and piecewise learning rate starting at 1 × 10 −6 , decaying at every five epochs (decay factor of 0.96) until it reached 1 × 10 −6 .
With the augmented images dataset, 30 original WPCB images were reserved for the test dataset, and the other images were split into training (80%) and validation (20%). The metric used to evaluate the IC detection was the all-point mean average precision (mAP) [24]. Since there was a single class, the mAP was the same as the AP (average precision) for the IC class, defined as , where R is a recall level and max : . In order to obtain the IC surface area, we needed to evaluate the dataset image scale. The silkscreen printing easily identified some IC part numbers on the package. With the measurement of the IC dimensions present in the datasheet, it is possible to calculate the real object size scale. It is important to note that the dataset authors do not provide information about the camera calibration. Table 2 shows the pixel area estimate for a particular IC (part number RTL810L). For each PCB, the dataset has images in different positions. This is important because usually, WPCBs are placed on conveyor belts with arbitrary orientation, and a computer-vision system has to perform visual inspection despite the PCB orientation. In order to obtain the IC surface area, we needed to evaluate the dataset image scale. The silkscreen printing easily identified some IC part numbers on the package. With the measurement of the IC dimensions present in the datasheet, it is possible to calculate the real object size scale. It is important to note that the dataset authors do not provide information about the camera calibration. Table 2 shows the pixel area estimate for a particular IC (part number RTL810L).  For all experiments, we used the pre-trained YOLOv3 provided by Darknet [16]. The fine-tune strategy was to restore the weights from all layers except the last three convolutional layers. Additionally, the YOLOv3 head was fine-tuned with a momentum optimizer and piecewise learning rate starting at 1 × 10 −6 , decaying at every five epochs (decay factor of 0.96) until it reached 1 × 10 −6 .

IC Object Detecion
With the augmented images dataset, 30 original WPCB images were reserved for the test dataset, and the other images were split into training (80%) and validation (20%). The metric used to evaluate the IC detection was the all-point mean average precision (mAP) [24]. Since there was a single class, the mAP was the same as the AP (average precision) for the IC class, defined as , where R is a recall level and max : .
(5) For all experiments, we used the pre-trained YOLOv3 provided by Darknet [16]. The fine-tune strategy was to restore the weights from all layers except the last three convolutional layers. Additionally, the YOLOv3 head was fine-tuned with a momentum optimizer and piecewise learning rate starting at 1 × 10 −6 , decaying at every five epochs (decay factor of 0.96) until it reached 1 × 10 −6 .

IC Object Detecion
With the augmented images dataset, 30 original WPCB images were reserved for the test dataset, and the other images were split into training (80%) and validation (20%). The metric used to evaluate the IC detection was the all-point mean average precision (mAP) [24]. Since there was a single class, the mAP was the same as the AP (average precision) for the IC class, defined as where R is a recall level and

IC Object Detecion
The achieved results for IC object detection are shown in Table 3. The precision was 0.917, and the mAP was 86.77%. Figure 6 shows a qualitative analysis of the results. It is possible to notice small deviations related to bounding boxes, in general, size, centralization, and rotation. However, even with these problems, a proper generalization of the trained model was noticed. For example, in some labels, specific components were not considered integrated circuits, but the model was able to detect them. The achieved results for IC object detection are shown in Table 3. The precision was 0.917, and the mAP was 86.77%. Figure 6 shows a qualitative analysis of the results. It is possible to notice small deviations related to bounding boxes, in general, size, centralization, and rotation. However, even with these problems, a proper generalization of the trained model was noticed. For example, in some labels, specific components were not considered integrated circuits, but the model was able to detect them.  Table 3. Precision, Recall, and mAP for the IC detector model.

IC Area Results
Using the pixel area scale, Figure 7 shows the predicted and ground truth IC area in 2 for each WPCB image. It is possible to note that the PCB DSLR Dataset had some images with high IC density (high IC area) and others with very low IC density (low IC area) or even WPCBs without any IC (IC area equaled zero).  Figure 8a shows the graph of the root mean squared error (RMSE) between the ground truth IC area and the YOLO predicted IC area. The highest RMSE value was 6192.38. The RMSE value had a high dispersion (average of 477.12 mm 2 and a standard deviation of 698.61 mm 2 ). However, most WPCBs had a percentage error below 2%, as seen in Figure 8b.

IC Area Results
Using the pixel area scale, Figure 7 shows the predicted and ground truth IC area in mm 2 for each WPCB image. It is possible to note that the PCB DSLR Dataset had some images with high IC density (high IC area) and others with very low IC density (low IC area) or even WPCBs without any IC (IC area equaled zero).
Appl. Sci. 2021, 11, x FOR PEER REVIEW 8 of 11 The achieved results for IC object detection are shown in Table 3. The precision was 0.917, and the mAP was 86.77%. Figure 6 shows a qualitative analysis of the results. It is possible to notice small deviations related to bounding boxes, in general, size, centralization, and rotation. However, even with these problems, a proper generalization of the trained model was noticed. For example, in some labels, specific components were not considered integrated circuits, but the model was able to detect them.

IC Area Results
Using the pixel area scale, Figure 7 shows the predicted and ground truth IC area in 2 for each WPCB image. It is possible to note that the PCB DSLR Dataset had some images with high IC density (high IC area) and others with very low IC density (low IC area) or even WPCBs without any IC (IC area equaled zero).

Estimated IC Weight
Using the calculated IC surface density, the total IC weight was calculated based on the total IC area. Table 4 shows the total area and weight for the ground truth and YOLO predictions. The error between the IC weighed calculated from the ground truth, and the one calculated from YOLO predictions was 160.72 g.

Calculated Recycling Metal Weight
Metals are commodities whose values fluctuate according to market factors. In order to calculate the financial return of each PCB, the average value of metal ore was collected from the stock market. Table 5 shows the estimated return of recycling of all the PCB's ICs, and the right axis of Figure 7 shows the return of IC recycling for each PCB. The maximum return was USD0.59 (USD0.69 with the IC detector model estimative), and the average return was USD0.05 (standard deviation of USD0.06). It is important to state that these incomes do not consider precious metals as gold and silver, so it is a fraction of the recycling outcomes.

Discussion
The proposed economic evaluation output was a total of USD7.44 (USD8.55 with IC detector model) with USD0.59 as the highest value for a PCB and (USD0.69 with IC detector model estimative), and the average return was USD0.05 (standard deviation of USD0.06). However, precious metals (Au, Pd, Pt, and Ag) are a small fraction of the ICs' mass and were not considered in this small sample. As precious metals are responsible

Estimated IC Weight
Using the calculated IC surface density, the total IC weight was calculated based on the total IC area. Table 4 shows the total area and weight for the ground truth and YOLO predictions. The error between the IC weighed calculated from the ground truth, and the one calculated from YOLO predictions was 160.72 g.

Calculated Recycling Metal Weight
Metals are commodities whose values fluctuate according to market factors. In order to calculate the financial return of each PCB, the average value of metal ore was collected from the stock market. Table 5 shows the estimated return of recycling of all the PCB's ICs, and the right axis of Figure 7 shows the return of IC recycling for each PCB. The maximum return was USD0.59 (USD0.69 with the IC detector model estimative), and the average return was USD0.05 (standard deviation of USD0.06). It is important to state that these incomes do not consider precious metals as gold and silver, so it is a fraction of the recycling outcomes.

Discussion
The proposed economic evaluation output was a total of USD7.44 (USD8.55 with IC detector model) with USD0.59 as the highest value for a PCB and (USD0.69 with IC detector model estimative), and the average return was USD0.05 (standard deviation of USD0.06). However, precious metals (Au, Pd, Pt, and Ag) are a small fraction of the ICs' mass and were not considered in this small sample. As precious metals are responsible for most of the financial return of WPCB recycling, an increase in the gross income of IC recycling is expected when also recovering precious metals like gold and silver.
The complete estimation of WPCB compositions depends on more data regarding the composition of each electronic component. Research in these directions could reduce the uncertainty in the evaluation of a WPCB composition.
A first application of the proposed method is to cluster PCBs with similar IC densities. Those similar WPCBs could suffer the same recycling process because they probably have similar compositions. Another critical component is the market value of the metal. The market value is positively affected by the purity of the metal. Therefore, depending on the used recycling technique, the gross outcome may be different.

Conclusions
Recycling waste printed circuit boards (WPCBs) is very important for recovering high-value metals and providing an environmentally safe destination for these electronic residues. Integrated circuit detection and other components can improve the PCB recycling process since it improves automatic disassembly systems and reuses some components. In this sense, the PCB DSLR public dataset was published to support the research and development of computer-vision strategies for PCB recycling. This paper proposed a WPCB economic feasibility assessment (WPCB-EFA) method for recycling ICs from PCBs. As WPCBs have a high composition diversity, the main goal is to provide information regarding the WPCB composition to guide the recycling process such as the need to dismantle a particular PCB before the physical or mechanical recycling process.
The first step of the WPCB-EFA is to detect ICs in a PCB. Deep learning object detection techniques require large-scale and fully-annotated datasets. Since the PCB DSLR dataset contains only 748 images from 165 different PCBs, transfer learning from the YOLOv3 pre-trained model was evaluated. In order to achieve better results in IC detection, the PCB DSLR Dataset images were split into squares of 416 × 416 pixels, resulting in a total of 5347 images containing integrated circuits. With this image size, YOLOv3 fine-tuning resulted in 0.965 mAP. Qualitative analysis suggests that the model's errors occur when part of the IC is in the image.
Considering the need for calculating the IC weight, we provide an IC surface density estimate based on the IC manufacturing data. Together with ICs recycling outcomes reported in the literature, we calculated the expected recovery mass for some elements. We noted that some PCBs have much more value than others and may go through different recycling processes. For instance, some PCBs may go to crushing, while others need dismantling before directly crushing. The proposed method may help to optimize the recycling process by such analysis.
The economic evaluation of PCB recycling can be one motivational factor for recycling, aside from the urgent need to deal with electronic waste. In future works, it is necessary to expand the recycling economic evaluation to other PCB components aside from integrated circuits.

Conflicts of Interest:
The authors declare no conflict of interest.