Multi-branch Recurrent Attention Convolutional Neural Network with Evidence Theory for Fine-Grained Image Classification

Fine-grained image classification (FGIC) aims to classify subordinate classes belonging to the same meta category. One of the existing FGIC methods is to use attention mechanism to localize and crop a discriminative region from the input image, and this process can be executed recurrently. In this way, the cropped image will progressively focus on a smaller local region containing the object part. However, this may cause the contour information of the object to be incomplete at the finest-scale and thereby the accuracy of the finest-scale is affected. In addition, the fusion strategy of these methods, which generally concatenates the outputs of multiple scales for the final classification, is not sufficient. To tackle the problems, based on the backbone of RA-CNN we first construct a multi-branch attention proposal network (APN) at middle scale of RA-CNN to jointly localize a most discriminative region where multiple APNs can complement each other’s incomplete contour information. Moreover, in addition to concatenating the outputs of all scales, we also use the Dempster’s combination rule to combine the outputs of all scales. Then, the features of these two parts are further combined for the final classification. Experimental results on the real-world datasets clearly validate the effectiveness of the proposed method.