YOLOv4 target detection algorithm for improving loss function

Abstract

The invention provides a YOLOv4 target detection algorithm for improving a loss function, a position regression loss function CIoULoss in a standard YOLOv4 model is improved, a novel loss function SCIoU is provided and embedded into YOLOv4, and performance improvement is achieved. The method comprises the following steps: firstly, downloading a general data set tt100k and LISA in a current target detection field, and performing data enhancement; secondly, training the two enhanced general data sets by using a standard YOLOv4 network and detecting the performance of the two enhanced general data sets; secondly, for a position regression loss function CIoULoss in the standard YOLOv4 model, providing an improved loss function SCIoU, and embedding the improved loss function SCIoU into the YOLOv4 model for training; and finally, comparing with a standard YOLOv4 algorithm, and analyzing a test result. The invention provides an improved YOLOv4 algorithm based on SCIoU, which comprises the following steps of: improving a length-width ratio measurement index vs, and replacing an arctan function with a Sigmoid function; according to the improved YOLOv4 algorithm, the algorithm is embedded into the YOLOv4, the performance is improved, no more calculation amount is introduced into the model, the real-time performance is not affected, and the improved YOLOv4 algorithm is good in robustness and can be used for improving the performance of multiple data sets.

Description

technical field [0001] The invention belongs to the field of image recognition, and in particular relates to a YOLOv4 target detection algorithm with an improved loss function, which shows good detection performance on common standard data sets. Background technique [0002] With the continuous development of computer technology, computer vision and target detection have become popular directions. Target detection can be used to identify and locate specific objects, and it has broad development prospects in driving assistance systems and military early warning systems. Target detection technology includes traditional target detection technology and deep learning-based target detection technology, and the latter has become the mainstream algorithm in the current target detection field because it is superior to the former in terms of performance and complexity. [0003] The target detection technology based on deep learning is mainly divided into two methods, one-stage and tw...

AI Technical Summary

Problems solved by technology

IoU_Loss uses the intersection-over-union ratio (IoU) as the loss function, but gradient return cannot be performed when the two frames do not intersect, and when the two frames overlap in different areas, the degree of overlap cannot be distinguished; in view of the above shortcomings, GIoU_Loss is proposed, and two frames are introduced The minimum cover area, but when the real frame completely contains the predicted frame, it will degenerate into IoU_Loss to slow down the convergence; in order to further improve the regression accuracy, DIoU_Loss is proposed, and the Euclidean distance between the center points of the two frames is introduced; on this basis, the predicted Considering the aspect ratio of the box to the real box, CIoU_Loss is proposed

However, when the CIoU_Loss loss function deals with small targets, the regression loss is also large when the target aspect ratio changes greatly, which is not conducive to the stability of model training.

Images