IEMAM: Image-Feature Extraction Based on Multi-Scale Attention Mechanism

Large variations in the scale of visual entities make it a challenge to reduce the computational complexity of image-feature extraction. To address this challenge, this paper presents an image-feature extraction method based on multi-scale attention mechanism(IEMAM). The method first introduces a hierarchical image shrinkage strategy, which uses linear projection to control the size of image features at each layer. Next, a multi-head attention module based on dimensional trimming is proposed to reduce computational complexity. Finally, IEMAM constructs an attention aggregation structure to aggregate each local feature of multi-scale images. These qualities of IEMAM make it compatible with a better performance of object detection tasks: on the MS COCO dataset, the average accuracy of AP, AP50, and AP75 is improved by 1.98%, 5.77%, and 3.45% compared to Pyramid Vision Transformer. On the dam defect image datasets, IEMAM improves 1.21% in mAP while significantly reducing computational overhead for the same input conditions.