面向密集场景结合TC-YOLOX的小目标检测方法

Small target detection method for dense scenes combined with TC-YOLOX

密集场景下小目标的高效精确检测是目标检测领域的关键问题。为了解决环境的多样性和小目标自身复杂性存在着特征难以提取、检测精度低等问题,提出一种面向密集场景结合TC-YOLOX的小目标检测方法。首先,通过在CSPNet中引入Transformer Encode模块,不断更新目标权重实现增强目标特征信息,提高网络的特征提取能力;其次,在特征金字塔网络中增加卷积注意力机制模块,关注重要特征并抑制不必要特征,提高不同尺度目标的检测准确度;然后,采用CIoU代替IoU作为回归损失函数,使得模型训练过程中网络收敛更快,性能更好;最后在PASCAL VOC 2007数据集上验证。实验结果表明,所设计的TC-YOLOX模型能够有效的检测出多样化场景中正常、密集、稀疏、黑暗条件下的小目标物体,mAP和检测速度可以达到94.6%和38 fps,与原始模型相比提升了10.9%和1 fps,对多种密集场景下的小目标检测任务均具有较好的适用性。

Efficient and accurate detection of small targets in dense scenes is a key problem in the field of target detection.In order to solve the problems of diversity of environments and complexity of small targets,such as difficult feature extraction and low detection accuracy,a small target detection method for dense scenes combined with TC-YOLOX is proposed.Firstly,by introducing Transformer Encode module into CSPNet,the target weight is continuously updated to enhance the target feature information and improve the network feature extraction capability.Secondly,the convolutional attention mechanism module is added to the feature pyramid network to focus on important features and suppress unnecessary features,so as to improve the detection accuracy of targets of different scales.Then,CIoU is used to replace IoU as the regression loss function,which makes the network converge faster and has better performance in the process of model training.Finally,it is verified on PASCAL VOC 2007 dataset.The experimental results show that the designed TC-YOLOX model can effectively detect small target objects under normal,dense,sparse and dark conditions in diversified scenes.The mAP and detection speed can reach 94.6%and 38 fps,which is 10.9%and 1 fps higher than the original model.It has good applicability to small target detection tasks in multiple dense scenes.