基于改进YOLOv5m的弱小目标识别方法
2022,30(12):218-223
摘要：针对空对地观测弱小目标识别与跟踪技术需求，提出了一种改进型YOLOv5m网络的多目标识别检测方法，以提升对所占像素个数小于10*10弱小目标的识别能力；分析了网络结构输入端Mosaic数据增强、Anchor计算、Focus模块及SPP模块对弱小目标的影响；在深度学习网络Prediction层引入距离交并比非极大值抑制（DIoU-NMS）代替传统非极大值抑制（NMS），引入距离交并比损失函数（DIoU_Loss）代替广义化交并比损失函数（GIoU_Loss)，加快边界框回归速率，提高定位精度，消除重叠检测，并在网络中引入4*4以上像素的目标识别层，提升对遮挡重叠弱小目标识别的准确率；实验结果表明，改进的深度学习网络算法与经典的YOLOv5m网络相比，目标识别的均值平均精度mAP指标达到89.7%，对比原网络提高了4.1%，实现了对图像像素个数小于10*10的弱小目标高精度识别，有效提升了深度学习网络对弱小目标的适应性和应用价值。
关键词：多目标识别；yolov5m；损失函数；目标识别层；深度学习
Weak and Small Targets Recognition Method Based on Improved YOLOv5m
Abstract：Aiming at the needs of air-to-earth observation weak and small target recognition and tracking technology, an improved multi-target recognition and detection method of YOLOv5m network is proposed to improve the recognition ability of weak and small targets with less than 10*10 pixels; The influence of Mosaic data enhancement, Anchor calculation, Focus module and SPP module on weak and small targets at the input end of the network structure is analyzed; In the Prediction layer of the deep learning network, the distance intersection over union non-maximum suppression (DIoU-NMS) is introduced to replace the traditional non-maximum suppression (NMS), and the distance intersection over union loss function (DIoU_Loss) is introduced to replace the generalized intersection over union loss function (GIoU_Loss), speed up the bounding box regression rate, improve the positioning accuracy, eliminate overlapping detection, and introduce a target recognition layer with more than 4*4 pixels in the network to improve the accuracy of occlusion overlapping weak and small targets; The experimental results show that, compared with the classic YOLOv5m network, the improved deep learning network algorithm achieves an average average precision mAP index of 89.7%, which is 4.1% higher than the original network, and realizes the image pixel number less than 10*10. The high-precision identification of weak and small targets effectively improves the adaptability and application value of the deep learning network to weak and small targets.
Key words：multi-target recognition; yolov5m; loss function; object recognition layer; deep learning
收稿日期：2022-05-05
基金项目：航空科学基金(202000190U1002)
