Abstract
Effectively disassembling and recovering materials from waste electrical and electronic equipment (WEEE) is a critical step in moving global supply chains from carbon-intensive, mined materials to recycled and renewable ones. Conventional recycling processes rely on shredding and sorting waste streams, but for WEEE—which is comprised of numerous dissimilar materials—targeted disassembly offers a superior path for improved material recovery. However, many WEEE objects share key external features and look nearly identical, despite varying internal component layouts and material compositions. This work introduces RGB-X, a multi-modal image classification approach that utilizes features from external RGB images and internal X-ray images. By employing a novel Iterative Class Activation Mapping (iCAM) architecture and a synthetic data generation pipeline using domain randomization, we achieve a classification accuracy of 98.6% on 10 generations of modern smartphones, significantly outperforming individual RGB (89.1%) and X-ray (97.9%) models.
Introduction
As the global community moves toward a circular economy, the recovery of precious materials—such as rare earth elements, tungsten, and steel—from electronic waste is paramount. Current industrial recycling often utilizes "shred and sort" methods, which are inherently inefficient and lead to material cross-contamination. Targeted disassembly, performed by specialized robotic systems, allows for cleaner material separation. The bottleneck for such systems is the initial classification step: a robot must know exactly what device it is handling to apply the correct disassembly sequence.
Identifying electronics is notoriously difficult. For example, an iPhone 7 and an iPhone 8 are visually indistinguishable to most standard computer vision systems, especially when the devices are damaged, dirty, or missing external components. While X-ray imaging can reveal internal layouts that distinguish these models, the lack of large, annotated X-ray datasets has historically hindered the development of robust deep learning models in this domain.
Methodology
Iterative Class Activation Mapping (iCAM)
To address the need for fine-grained classification, we developed the iCAM architecture. Rather than processing an entire image in a single pass, iCAM utilizes an attention-based iterative strategy. The model identifies the most discriminative region for classification (the "heat map"), crops that specific area, and passes it back through the network. By performing this iteratively (typically $n=3$), the model "zooms in" on the subtle details—such as a specific battery connector or a unique screw pattern—that differentiate similar models.
Synthetic X-ray Data and Domain Randomization
To overcome the scarcity of labeled X-ray data, we developed a novel synthetic data pipeline. We utilized 3D CT scans of electronic devices and projected them into 2D space to simulate X-ray imagery. To ensure these synthetic images generalized to real-world scans, we applied Domain Randomization (DR). This involved randomly varying parameters such as device orientation, sensor noise, and "transfer functions" (which simulate different X-ray machine energy settings). This approach allowed us to train models on approximately 25,000 synthetic images that translated seamlessly to real-world hardware.
RGB-X Fusion
The RGB-X system processes color images and X-ray images through separate iCAM streams. The features extracted from both modalities are then integrated using a "Fast Fusion" method in the deeper layers of the network. This allows the model to rely on external visual cues (like logos or port placements) when internal data is noisy, and internal layouts when the external casing is deceptive or damaged.
Experimental Results
Performance on Smartphone Datasets
We evaluated our approach on a dataset of 10 generations of modern smartphones (SP1-SP10). The results demonstrate the clear advantage of multi-modal sensing:
- RGB Only: 89.1% accuracy
- X-ray Only: 97.9% accuracy
- RGB-X Fusion: 98.6% accuracy
The high performance of the X-ray-only model, trained entirely on synthetic data, highlights the effectiveness of our domain randomization pipeline. Ablation studies showed that varying the 3D projector position was the most critical factor; without it, accuracy dropped to 31.5%.
Generalization and Speed
To verify the iCAM architecture's versatility, we tested it on the CUB-200 bird species dataset, where it proved effective for general fine-grained classification tasks. In terms of industrial utility, the RGB-X model operates at approximately 30 FPS, making it suitable for real-time sorting on high-speed industrial conveyor belts.
Conclusion and Future Work
RGB-X provides a robust framework for the automated identification of complex electronic waste. By combining the external context of RGB with the internal structural data of X-rays, we can achieve the precision required for automated disassembly. Future research will focus on expanding the dataset to include a wider variety of manufacturers and evolving the iCAM module to attend to multiple regions of interest simultaneously. Furthermore, we aim to improve the interpretability of these models to provide clear justifications for classification decisions in industrial settings.
Summary updated