FINE-GRAINED VISUAL CLASSIFICATION BASED ON ADAPTIVE STRIDE CONVOLUTION

Down-sampling methods such as average pooling have been widely used to reduce computation cost, prevent overfitting, and improve the performance of convolutional neural networks. However, in fine-grained recognition tasks, these uniform sampling methods cannot focus well on subtle discriminative regions. In this paper, we propose an Adaptive Stride Convolution Network (ASCNet) in which the ASC module is used to focus on extracting subtle features. Specifically, given an image, we obtained an attention map to highlight the discriminative parts of object, where the attention map extractor was used. The attention map-based stride generator produced stride vectors which indicated the moving steps of convolutional kernels every time. The adaptive stride convolution extracted information over the input image or features with varying strides. We experimentally evaluated the effectiveness of our method on three challenging JP3fine-grained benchmarks, i.e.KG-*3, CUB-200-2011, Stanford Cars, and FGVC-Aircraft, and advanced performance is achieved.