Abstract: Remote sensing images often contain numerous objects. Since it is challenging to extract effective category discrimination information, the existing remote sensing object detection methods suffer from categories confusion phenomenon. Therefore, a discriminative region perception enhancement method has been proposed for fine-grained remote sensing object detection. First, a discriminative information extraction module is constructed within the feature extraction backbone network. With the help of discriminative feature enhancement and suppression strategies, local details information of objects that are easily overlooked could be continuously extracted, the features representation power of the network for different types of objects are enhanced. Second, an object perception enhancement module is built, which utilizes the feature selection module to enhance the multi-scale feature description power. Besides, multiple convolution layers with different kernel sizes and skip connections are added to improve the object awareness of shallow layer. Finally, inspired by self-attention mechanism in transformer network, a self-attention feature enhancement module is constructed to highlight the features of detected objects and reduce interference of complex backgrounds on the accuracy of object classification. The experiments on the fine-grained remote sensing object detection datasets DOSR and HRSC2016 indicate that the proposed method achieves mean average precision of 72.04% and 84.38%. Qualitative and quantitative experiments indicate that the proposed method exhibits superior accuracy and robustness in detecting various categories of remote sensing objects.