Normal Distribution Sampling Convolutional Neural Network for Fine-Grained Image Classification

In this paper, we propose a Normal distribution Sampling Convolutional Neural Network (NS-CNN) for Fine-grained Image Classification. Different from other fine-grained classification networks, which directly use the discriminative feature area for classification, NS-CNN centers on the discriminative feature area, and uses the probability model based on the two-dimensional normal distribution to resample the image pixels, and then recognizes the new image. This method can focus on the discriminative feature area and consider the influence of the surrounding areas of the discriminative feature. At the stage of classification, NS-CNN divides the image into several grids, and each grid is classified as a discriminative feature area. And finally, all the classification results are merged to get the final result. This method omits the complex discriminative feature localization network, so the network parameters are fewer. This helps enable fine-grained classification networks to work on computers with common hardware. We tested this method on CUB-200-2011 dataset, and the experimental result show that NS-CNN can get an outstanding performance on fine-grained classification with a lightweight network architecture.