Classification of imbalanced municipal solid waste image datasets

To address model overfitting and long-tail omissions caused by the shortage of minority class samples in municipal solid waste image collection, this study proposed a classification algorithm integrating an Improved Generative Adversarial Network, a Multi-Scale Attention Network, and a Dynamic Focal Loss. The proposed method employed an Improved Generative Adversarial Network to alleviate data imbalance by expanding minority class images, a Multi-Scale Attention Network to capture multi-scale spatial semantics, and a Dynamic Focal Loss to adaptively adjust gradient weights for difficult samples. Experimental results reveal an accuracy of 95.88%, a recall of 96.14%, and a misclassification ratio for minority categories of less than 0.03%. For simulated experiments, the inference latency is reduced to 14.8 ms, throughput achieved is 67.5 FPS, and the memory requirement is approximately 1.8 GB. Based on these findings, it can be concluded that the proposed approach demonstrates great anti-confusion and lightweight nature in extremely imbalanced data distribution scenarios. This study provided a reliable solution for industrial-grade intelligent waste sorting.