Hierarchical feature attention with bottleneck attention modules for multi-branch classification

Abstract

While existing attention mechanisms often focus on pre-processing images, fine-grained classification tasks benefit from leveraging hierarchical relationships within categories. For example, classifying bird species involves understanding broader categories like orders and families. This inherent structure helps reduce ambiguity in predictions. This work proposes a novel approach that integrates Bottleneck Attention Mechanisms (BAM) within a ResNet50 backbone for multi-task classification. By employing separate feature branches tailored to each task and applying BAM after each branch, the model learns more discriminative features for each hierarchy. This report details the architecture and training strategy of this proposed model.