基于特征增强和分组模块的车型精细识别

Fine Vehicle Recognition Based on Feature Enhancement and Grouping Module

针对车型种类多、差异小,模型复杂,识别精度低的问题,提出一种基于特征增强和分组模块的车型精细识别方法,在ResNet网络基础上改进,在卷积块中加入多尺度通道域和空间域的注意力机制,增强对重要的特征提取,并将多通道特征图进行分组,根据分组损失函数不断优化分组,通过加权方式结合KL(Kullback-Leibler)散度损失函数和交叉熵损失函数,有助于网络学习类内差异小、类间差异大的特征。该方法在Stanford cars-196数据集和自制数据集上进行测试,验证了所提模型的有效性。

In order to solve the problems of many types of vehicles,small differences,complex models and low recognition ac-curacy,a fine vehicle recognition method based on feature enhancement and grouping module is proposed,which is improved on the basis of ResNet network.The attention mechanism of multi-scale channel domain and spatial domain is added to the convolution block to enhance the extraction of important features,and the multi-channel feature graphs are grouped and continuously optimized according to the grouping loss function.KL(Kullback-Leibler)divergence loss function and cross entropy loss function are com-bined by weighted method.The method is tested on Stanford cars-196 dataset and self-made dataset to verify the effectiveness of the proposed model.