From Fuzzy Global to Clear Local: A Focus and Super-Resolution-Guided Tiny Target Detection Method for Full-Scene Images

Detecting tiny targets, a fundamental satellite observation task, always suffers from information loss, additional computation costs, and imbalance samples in full-scene imagery. However, generic methods rely on complex deep neural networks to traverse every region of the image, irrespective of the non-uniform target distribution, severely hindering the detecting accuracy and leading to a heavy computation burden. This paper proposes a feature focusing super-resolution small target detection network (FFSSTD-Net), alleviating the above three problems. First, the proposed convolution focus detection module (CFD) uses a lightweight full-convolutional network query mechanism to adaptively search the regions with targets, eliminating unnecessary computation costs and noise from the background. Then, the proposed feature super-resolution (FSR) module guides the backbone network enhancing the representation of small target objects in high-dimensional features through an auxiliary branch to achieve a balanced optimization, alleviating the information loss caused by sub-sampling operations. Experimental results demonstrate that the FFSSTD-Net outperforms state-of-the-art generic methods, improving mAP and Fps by at least 3.14% and 10%, respectively. Meanwhile, ablation experiments indicate that the CFD and FSR modules enhance detection accuracy while accelerating full-scene image processing.