<![CDATA[Waste sorting automation through deep learning is an important approach to support sustainable waste management systems. However, many studies still overlook dataset quality issues, especially exact and near duplicate images that can cause information leakage and overly optimistic evaluation metrics. This study proposes a waste image classification pipeline that integrates an explicit data quality analysis stage using MD5 hashing for exact duplicate detection and perceptual hashing (pHash) for near duplicate detection, followed by fine tuned EfficientNet B4 as the classification backbone. Experiments are conducted on two public datasets with distinct characteristics: Garbage Classification V2 (6 classes, 9,421 images) and RealWaste (9 classes, 4,749 images). With a Hamming distance threshold τ≤2, the pHash cleansing identifies zero duplicates in Dataset 1 and only three near duplicates (0.06%) out of 1,404,077 compared pairs in Dataset 2, confirming no evidence of image-duplication-based information leakage in either dataset. EfficientNet B4 achieves 97.77% test accuracy with a macro F1 Score of 0.9768 on Dataset 1 and 93.26% accuracy with a macro F1 Score of 0.9379 on Dataset 2, demonstrating consistent performance across these two datasets with different numbers of classes, data volumes, and visual heterogeneity. These findings should be interpreted as evidence of robustness within the scope of the two evaluated datasets, rather than as a claim of generalization to unseen, external, or cross-domain waste image datasets.]]>
