Heatmap Visualization for Deep Learning Analysis of Waste Printed Circuit Boards

Waste Printed Circuit Boards (WPCBs) are complex multi-material assemblies that present challenges for automated recycling and Critical Raw Material (CRMs) recovery. Visualization of the part of the WPCBs need more attention and contain high-level density CRMs is challenging in computer vision based system analysis. In this work, we propose a deep learning-based multi-label classification framework integrated with heatmap visualization for interpretable WPCB analysis. We fine-tuned the ResNet50 model as backbone and applied binary cross entropy for each class on custom multi-label V-PCB dataset converted from YOLO format. For visualization of the specific regions across the WPCBs with an image, we applied Gradient-weighted Class Activation Mapping (Grad-CAM) that generate class-specific activation maps corresponding to high density CRMs contained components. Experiments on a custom curated V-PCBs dataset achieve a micro-averaged F1 score of 97.67%. The proposed system provides accurate classification along with interpretable heatmaps, supporting automating vision-based disassembly methods and recovery processes in e-waste recycling.