Pixel Attention Feature Pyramid Network for Few-Shot Object Detection in Remote Sensing Images

Recently, few-shot object detection (FSOD) has attracted extensive attention in the field of computer vision. However, these FSOD methods for natural images have limited performance when applied to remote sensing images with arbitrary orientations, dense objects, and complex backgrounds. In this paper, we propose a novel few-shot object detection method to address the above challenges in remote sensing images. First, we design a non-freezing two-stage fine-tuning detector that is fully trained on a base class data to extract information in the first stage, and fine-tunes the entire network for better novel class detection in the second stage using balanced training data containing both base and novel class samples. Then, a pixel attention feature pyramid network (PA-FPN) is proposed to replace the traditional FPN, which adds a pixel attention-based feature selection module (PA-FSM) and a pixel attention-based sub-pixel convolution module(PA-SPCM). In PA-FSM, pixel attention is used to extract more discriminative channels, thereby highlighting key features while suppressing background noise. In PA-SPCM, the sub-pixel operation performs convolution and upsamples the low-resolution feature maps with the aim of obtaining a high-resolution feature maps with richer semantic information, while pixel attention is applied to the high-resolution feature maps to enhance the semantic information. Finally, we introduce a proposal rectification head to eliminate false positives with high scores and address missing samples with low scores, which can improve classification accuracy. Experiments conducted on two public remote sensing datasets demonstrate the effectiveness of our method.