基于局部特征融合的细粒度车辆识别

Fine-grained vehicle recognition based on fusion of part feature

为有效提高基于局部检测的细粒度图像分类方法的工作效率,提出一个自适应通道分配模块,能主动分组表达相同语义信息的特征通道。此过程的学习由设计的判别性和多样性损失函数监督完成,利用多尺度深度可分离卷积,从已提取的全局图像特征中检测有助于分类的多样化局部信息。通过训练的网络具有强大的特征分配能力,在全局对象定位的基础上进一步实现对细节的定位。图像的融合表示综合考虑各个部分对分类的贡献,有效分类细粒度车型,在公开的Stanford Cars和CompCars数据集上的对比实验结果验证了该方法表现良好。

To improve the efficiency of part-based fine-grained classification methods,an adaptive channel assembling model was proposed,which could independently grouped feature channels that expressed same part semantics.This learning process was completed under the supervision of designed discriminality and diversity losses by using multi-scale depthwise separable convolution to detect diversified parts conducive to classification from extracted global features.The trained network had strong channel assembling ability,so that it could further realized the localization of parts on the basis of object localization.The fusion description of images comprehensively considered the contribution of each part to final classification and effectively recognized fine-grained vehicle models.The comparative results on the published Stanford Cars and CompCars data sets verify that this approach performs well.