The main target detector happened to be using yolov5 recently. Say something personal, not necessarily right.

First of all, the yolo series has developed to the present, and the thinking is very mature. Like the specific improvement, in fact, 4 and 5 have many similarities, whether it is the spine or the neck.

At present, many papers about yolov5 have been improved, some are lightweight on backone, some are added with attention mechanism, some are improved with neck or head, and some are changed with loss function or nms process.

Mainly because yolo's own ideas are very mature, it is really difficult to make some universal innovations and upgrades under this framework. As for putting all kinds of mature modules into it to send papers, this kind of benevolent person has different opinions. Personally, I feel that improving yolo should be combined with a certain direction, and there will be some directions. After all, the same promotion sometimes has different effects on different data sets and scales.

The loss function of Yolov5 target detection consists of three parts, namely, rectangular frame prediction loss function, confidence prediction loss function and category prediction loss function. In the last section, the defects of the target detection loss function GIoU and its improvement are analyzed, and the loss function of the original network is replaced by CIoU and binary cross entropy function with adjustment factor.

Experiments verify that this improvement is compared with the original algorithm, as shown in the following table. According to the data in the above table, it can be seen that the improvement of the loss function in this paper has achieved an accuracy of 92. 1% in the experiment, which has increased by 0.5%, which can prove that the improvement of the loss function is very helpful to improve the target detection performance.