To enrich the representation information of the fine-grained image features of CNN,and expand the inter class diffe-rences while reduce the intra class differences,a feature fusion method based on the multi-granularity image training was pro-posed to excavate the thinning features of images.By constantly changing the granularity value of the input images,a model con-tained the multi-granularity information was constructed.And the multi-scale-granularity image features were extracted to fuse with the original features to complete classification.At the same time,it completed the question of fusing different granularity features without increasing the parameters significantly and introducing auxiliary networks.Experimental results show that the classification accuracy is higher than the results containing a single granularity image,which verifies that the model can effectively enrich the feature information.
维普
万方数据