Combining Local Detail Sensitivity with Global-Weighted Aggregation in Trash Classification

Trash image classification plays a crucial role in waste management and environmental sustainability, yet existing methods often struggle with intricate details due to their reliance on single-scale feature extraction techniques. This paper presents a novel approach that addresses these limitations by integrating multi-scale local feature enhancement with global-weighted feature aggregation. By employing diverse convolution kernels of varying sizes, the model enhances its sensitivity to local patterns within images, capturing more relevant features. Furthermore, the extracted feature maps are efficiently condensed into simplified vectors through global average pooling, while an information entropy-based weighting mechanism assigns significance to each vector based on its corresponding feature map’s entropy. This facilitates a nuanced relationship assessment between point features and global representation via normalized cosine similarity. Finally, our designed loss function combines cross-entropy loss with regularization to optimize performance while mitigating overfitting risks. Extensive experiments on public trash datasets demonstrate that this proposed method achieves state-of-the-art performance for this classification task.