Optimizing Receptive Field and Multi-Scale Features for Object Detection with Lightweight Hybrid Attention

Currently, object detection models have achieved a relatively ideal balance between speed and performance. However, taking the YOLO model as an example, it relies on Convolutional Neural Networks (CNN) to extract image features. Traditional convolutions face limitations in capturing multi-scale features, and increasing the receptive field often comes with significant computational overhead. Moreover, existing channel attention mechanisms tend to overlook the multi-scale feature space, while hybrid spatial-channel attention approaches face the challenge of high computational complexity. To address these issues, we propose a lightweight attention module that combines spatial and channel information, integrating both local and global features. Additionally, we design a novel multi-scale hybrid convolution module, which significantly enhances multi-scale feature representation and receptive field, while maintaining low computational complexity. We evaluated our method on the COCO and VisDrone datasets, and the experimental results demonstrate that the proposed model performs excellently across multiple metrics, significantly improving detection performance.