Fine-grained visual classification aims to recognise objects from subordinate categories. Although deep neural networks have achieved success in learning local differences between different objects, how to filter out meaningless feature channels to enhance the localized ability of models in deep neural networks remains under-explored in fine-grained classification scenario. In this work, we propose a Filtering and Localizing Network, namely FLNet, to achieve high-frequency feature channels mining and accurate targets’ localization under non-deformation object condition. The FLNet consists of two modules and a strategy: 1) High-frequency Channel Selection (HCS) module, 2) Localizing Discriminative Region (LDR) module, and 3) Non-deformable Object Region Cropping Strategy. By designing HCS, model can select high-frequency discriminative feature channels and filter out meaningless feature channels. Further, with the aid of HCS, the LDR can fully exploit channel dependencies, strengthening the model’s ability to locate object regions. And, the above two modules are done based on the Non-deformable Object Region Cropping Strategy, which can deal with the issue of objects’ deformation, boosting models’ performance on fine-grained visual classification. Extensive experiments conducted on four public fine-grained benchmarks demonstrate that FLNet can effectively ignore complicated backgrounds and localize target regions conducive to classifying fine-grained objects, achieving state-of-the-art performances.