Small targets detection in low-resolution remote sensing images based on super-resolution joint optimization

While convolutional neural networks have driven remarkable progress in remote sensing object detection, persistent challenges remain in detecting small targets within low-resolution imagery due to their limited pixel representation and feature degradation during hierarchical downsampling. To address this, this study proposed the joint super-resolution and detection network (JSRDN), which synergistically optimizes SR reconstruction through task-specific detection feedback, significantly enhancing small target recognition in LR remote sensing imagery. Firstly, generator in generative adversarial network incorporates improved residual blocks, enabling enhanced perception of complex deep-level features in the SR reconstruction process. Then, a perceptual loss function is introduced into the adversarial training process, which captures perceptual discrepancies in high-level features between reconstructed images and original HR references. After that, an edge-enhancement network is designed to dynamically detect edges in intermediate features restored by the generator, prioritizing edge influence across network layers to generate discriminative features for target recognition. Furthermore, the JSRDN implements detection-driven feedback by backpropagating object recognition loss through the generator, enforcing the super-resolution process to prioritize detection-salient feature recovery. Evaluated on 64×64 low-resolution COWC datasets, JSRDN achieves 0.1819 dB peak signal-to-noise ratio (PSNR) and 7.18 % average precision (AP) improvements over the deep residual dual-attention network (DRDAN), with ablation studies and visualizations confirming its balanced optimization of reconstruction fidelity and detection-oriented feature learning. This technology can provides valuable support for small target measurement and opens new opportunities in the field.