Distinguishing Glass and Plastic Bottles Using EfficientNet-B0 with Channel Attention and Caustic Pattern Analysis

With the increasing visual similarity between modern plastic and glass bottles—driven by advances in material processing technologies—accurate classification of these materials has become a growing challenge in sustainable waste management. Traditional image-based classification methods, even those using deep learning, often struggle to distinguish between plastic and glass due to overlapping visual features. To address this issue, we propose a novel classification approach leveraging caustic images—photographs capturing refractive light patterns produced as light passes through transparent materials—and an EfficientNet-B0 model enhanced with a channel attention mechanism. This approach capitalizes on the inherent differences in optical properties between glass and plastic, enabling the model to focus on texture and light scattering rather than color or shape. A dataset of approximately 13,000 caustic images was used, with extensive data augmentation applied to improve generalization under varying lighting conditions. Experimental results demonstrate the effectiveness of the proposed method, achieving a classification accuracy, precision, and recall of 99.9%. This study offers a robust, efficient solution for plastic-glass classification and contributes to the development of advanced, real-time waste-sorting technologies.