A Novel Object Detection Method Based on Adaptive Attention Mechanism and Composite Confidence

Currently, some object detection methods combine anchor-free frameworks with single-stage approaches, allowing the model to be freed from the computational burden of using anchor boxes. This provides more space and flexibility, bringing both detection and segmentation closer to real-time accuracy. However, anchor-free object detection algorithms are confronted with some deficiencies due to their immature development, particularly in multi-scale, sampling, and loss functions. To dispose of the aforementioned issues, an object detection method based on adaptive attention mechanism and composite confidence is proposed. First, a matrix feature network based on the adaptive attention mechanism is proposed as the detection neck framework to address the detection difficulties caused by targets of different scales and aspect ratios. Moreover, the adaptive attention mechanism enhances features and facilitates feature transmission and fusion, thereby improving the precision of feature selection. Then, an object detection method guided by composite confidence is raised to refine the non-maximum suppression process, effectively suppressing low-quality bounding boxes. Additionally, the regression loss function is improved to optimize gradients, thereby enhancing detection accuracy. Furthermore, an adaptive center-covered sampling strategy is presented, which improves the overall sample quality by filtering out edge sample points during the sampling process. To validate the availability of our proposed methods, extensive experiments are conducted on two standard datasets: MS COCO 2017 and PASCAL VOC 2007. Experimental results demonstrate that our method can achieve better accuracy and possess higher computational efficiency.