End-to-End Target Detection Enhanced by Hierarchical Convolution Modulation and Sparse Attention Transmission

This study proposes an end-to-end target detection model that significantly improves accuracy, particularly for both large and small targets. The key innovations lie in the hierarchical convolution modulation of deformable attention and a sparse feature transmission strategy. By modulating deformable attention with convolution, our approach enables effective extraction of both global and local features. Additionally, a deep-shallow layered convolutional module is employed to enhance coarse- and fine-grained feature extraction. The sparse transmission strategy ensures optimal feature output from the encoder layers. Experimental results on the MS-COCO dataset demonstrate the effectiveness of our approach, achieving state-of-the-art performance with notable gains in detection accuracy for both small and large targets.