A sparse focus framework for visual fine-grained classification

Abstract

The location of discriminative features and reduction of model complexity are the two main research directions in fine-grained image classification. The manual annotation of object is very labor-intensive, and the commonly used model compression methods usually reduce the classification accuracy while compressing the model. In this paper, we propose a Sparse Focus Framework(SFF) based on Bilinear Convolutional Neural Network(BCNN), which includes self-focus module and sparse scaling factors. The focus module like the focus function of human beings automatically locates the object from background without manual labeling, and only a small amount of computing resources are occupied in the guaranteed accuracy. The sparse scaling factor for each channel can evaluate the importance of feature channels, which is adopted in pruning of channels. The large number of parameters and calculations in the fine-grained classification model can been effectively reduced by pruning method adopted in our network, which can obtain a sparse structure to prevent overfitting while maintaining classification performance. Our experimental results show that our model obtains accuracy of 90.2%, 84.5% and 92.0% on FGVC-aircraft, Stanford dogs and Stanford cars, respectively. Compared with the highest classification accuracy obtained by the same classification network B-CNN[D,D], the accuracy gains with 6.1%, 4.1% and 1.4% respectively. Moreover, the channel-level sparsity effectively reduces 30% of the network parameters and nearly 13% of the computation.