基于YOLOX-s的改进研究

Research on Improvement Based on YOLOX-s

摘要: 随着深度学习的飞速发展，行为识别在计算机视觉中的应用也越来越多。作为行为识别的重要部分，目标检测算法的研究愈显突出。本文针对单阶段检测算法中效果较好的YOLOX-s模型，提出了基于YOLOX-s的注意力改进，在网络结构中，增加了SE模块和CBAM模块，将注意力集中在图片检测的目标区域中，使得更多的细节信息获取出来；同时选用DIoU损失函数，让聚集的目标框和检测框之间重叠部分更快地重合，提高了模型的精度。实验结果表明，所提的模型相比于YOLOX-s模型，在准确率、召回率以及mAP上有所提升，对满足实际需求更近一步。

Abstract: With the rapid development of deep learning, behavior recognition is more and more used in computer vision. As an important part of behavior recognition, the research of target detection algorithm is more and more prominent. Aiming at the YOLOX-s model with better effect in the single-stage detection algorithm, this paper proposes an attention improvement based on YOLOX-s. In the network structure, SE module and CBAM module are added to focus on the target area of image detection, so that more detailed information can be extracted; at the same time, DIoU loss function is selected to make the overlapping parts between the aggregated target frame and the detection frame coincide faster, which improves the accuracy of the model. The experimental results show that compared with the YOLOX-s model, the proposed model improves the accuracy, recall and map, and is a step closer to meeting the actual needs.