Automated E-Waste Sorting Using Swin Transformer for Efficient Integrated Circuit Package Classification

The accelerating accumulation of electronic waste (e-waste), particularly from semiconductor industries, presents growing environmental and operational challenges. One of the largest constituents of this electronic waste (e-waste) are integrated circuit (IC) packages, and these must be sorted effectively and cost-efficiently so that they can be efficiently recycled. Manual sorting methods, as yet widespread, are increasingly becoming inadequate with their manpower requirement, rate of errors, and inability to scale as e-waste grows larger. In this paper, we propose an improved artificial intelligence (AI)-based system to automatically detect and classify dismantled Surface-Mount Devices (SMD) and Through-Hole Packages (THP) from input images. The proposed system employs the Swin Transformer as a feature extraction backbone accompanied by a Feature Pyramid Network (FPN) for multi-scale detection enhancement and a DyHead module for precision enhancement with adaptive attention. The model is trained on a custom IC dataset in COCO format and fine-tuned using Adaptive Training Sample Selection (ATSS), which allows it to effectively learn from realworld complex data. The system achieved a mean Average Precision (mAP) of $\mathbf{6 8. 1 \%}$ and an Average Precision (AP) of 82.9% at an Intersection over Union (IoU) threshold of $\mathbf{0. 5 0}$. These results substantiate the effectiveness of the system in correctly identifying and classifying IC packages in different conditions. The proposed system is a step towards sustainable management of e-waste by reducing reliance on human sorting and establishing the impact of deep learning on industrial-scale processing of e-waste.