Multi-scale target detection network based on Libra R-CNN and traffic sign detection method

Abstract

The invention provides a multi-scale target detection network based on a Libra R-CNN. The network aims at solving the problem that traffic sign detection under urban roads is complex in environment. The number of types of targets is large and the number is unbalanced, the Libra R-CNN is improved, the method comprises the following steps of: modifying an IoU-based Sampling module; using the GA-RPNfor replacing the original RPN, and replacing the Balanced L1 Loss with the Smarth L1 Loss, so that more accurate and more diversified samples are generated in the training period, the detection accuracy is improved, and the effectiveness of the method is verified through experiments. The experiment is carried out on an MS COCO 2017 and a traffic sign data set. The mAP of the improved Libra R-CNNis improved by 3 percent points, and the mAP reaches 0.773. Experimental results show that compared with an original target detection network, the performance of the improved network is remarkably improved.

Description

technical field [0001] The invention belongs to the technical field of image processing, and in particular relates to a Libra R-CNN-based multi-scale target detection network and a traffic sign detection method based thereon, which can be used to detect traffic signs in urban road scenes. Background technique [0002] Machine vision and deep learning are now widely used in many applications. Relatively mature technologies have been developed. Large and medium-scale target detection has reached a high level of accuracy, but there are still many difficulties. Small target detection in complex backgrounds The problem and the data imbalance problem that commonly exists in multi-category label targets are two outstanding difficulties. They cause weight imbalance and seriously affect the accuracy of detection, but they have not been well solved in the world, and therefore It has become a key issue affecting the practical application of artificial intelligence. Traffic sign detect...

AI Technical Summary

Problems solved by technology

[0011] In the process of using IoU-balanced Sampling, there are still three problems that affect the training effect: (1) Due to the large number of background extraction frames (negative samples) and too few extraction frames (positive samples) containing targets, it is not considered The balance between positive and negative samples; (2) Although the processing of negative samples increases the number of difficult negative samples, the number of difficult negative samples is much smaller than that of simple negative samples, so the imbalance between difficult samples and simple negative samples is still exists; (3) the anchor box is a box that locates the target area

Images