Vehicle type classification network training method based on destructive learning and attention mechanism

Abstract

The invention relates to a vehicle type classification network training method based on a destructive learning and attention mechanism, which specifically comprises the following steps of S1, obtaining an initial image of a to-be-tested vehicle, and performing multi-scale cutting on the initial image according to a damage mechanism to form sub-images corresponding to multiple scales; S2, extracting a feature map of the sub-image according to the convolution layer; S3, carrying out channel splicing on the feature map of the small-scale sub-image and the feature map of the large-scale sub-image after carrying out up-sampling on the feature map of the small-scale sub-image to form an initial feature map layer, and sending the initial feature map layer to a channel attention module; S4, processing the initial feature map layer in a channel attention module to obtain a channel weight matrix, and performing calculation according to the channel weight matrix to obtain weighted feature maps of the sub-images; and S5, calculating the weighted feature map of the multi-scale sub-image according to a softmax regression function to obtain a target image category. Compared with the prior art, the invention has the advantages that the classification precision and classification accuracy of the classification network are improved, and the method is more suitable for vehicle type classification project requirements.

Description

technical field [0001] The invention relates to the technical field of image processing, in particular to a vehicle classification network training method based on destructive learning and attention mechanism. Background technique [0002] Since each vehicle brand and different sub-series will launch new vehicles every year, the diversification of vehicle models makes it more difficult to determine the modified vehicles during the annual vehicle inspection process. Therefore, the problem of vehicle classification has always been one of the difficulties in engineering practice. [0003] In engineering practice, by comparing the style pictures of vehicles of different brands, different sub-series and different ages, it is found that the overall appearance of vehicles of different brands may be different, but the overall appearance of vehicles of the same brand and different sub-series may be very similar, and the difference is only reflected in the vehicle Headlights, grilles,...

AI Technical Summary

Problems solved by technology

Therefore, when training the classification model of this kind of data set with high similarity of each category and the difference between categories mostly appears on some detailed features, if the traditional convolutional network combined with the softmax classification network is used to train it, at the same time the whole The image is used as the input image and the feature map is extracted. What the network learns more may be the difference in the overall appearance or texture or color of each category, and it is easy to ignore the difference in some detailed features.

Images