Data enhancement method for traffic sign target detection field

Abstract

The invention discloses a data enhancement method for the traffic sign target detection field. According to the method, a small number of categories in a traffic sign data set are cascaded with a plurality of data enhancement methods in a targeted manner for data enhancement, and the long tail problem of category distribution in the data set is solved; and meanwhile, the trained model can cope with complex actual conditions in a road scene. Different data enhancement strategies are used for each category, so that the image generated by the data enhancement algorithm does not damage the important characteristics of the traffic sign while enriching the characteristics of the traffic sign sample, and the accuracy of the model in various complex application scenes is improved; and an optimal data enhancement hyper-parameter group is searched by using a simulated annealing algorithm, and an optimal hyper-parameter solution is searched in the vast hyper-parameter space.

Description

technical field [0001] The invention relates to the technical fields of computer vision and target detection, in particular to a data enhancement method oriented to the field of traffic sign target detection. Background technique [0002] In the field of computer vision target detection, the target detection algorithm based on deep learning obtains the ability to infer the position and category of the target in the picture through the learning of the data set. The latest deep learning-based target detection models such as YOLOv5, Mask RCNN and other algorithms perform well on many data sets, but in order to prevent overfitting during the training process, a large amount of data is required for training to exert the effect of the algorithm. Due to the different occurrence probabilities of different types of traffic signs in the real world, many traffic sign datasets have serious sample imbalance problems, and the model cannot achieve high accuracy on traffic sign categories w...

AI Technical Summary

Problems solved by technology

These data enhancement methods can increase the sample size for pedestrians, vehicles and other targets, but for traffic signs, their color and shape carry very important feature information, and inappropriate data enhancement strategies will destroy their features and cause the model to fail during training. Underfitting occurs in the middle, for example, if the left turn sign is flipped horizontally, it will be confused with the right turn sign

At the same time, the position distribution of traffic signs in the picture is also an important feature. Using inappropriate data enhancement methods to make traffic signs appear where they should not appear will also reduce the accuracy of the model.

Images