Region-guided network with visual cues correction for infrared small target detection

Infrared small target detection (IRSTD) has experienced fast developments in recent years and been widely applied in civilian and military fields. The long imaging distance and complex backgrounds of infrared images often make the interested targets present in small scales and lack of contour features, which poses great challenges for the detection. Though deep neural network-based methods have been thoroughly investigated in IRSTD, deep layers generally struggle to retain the visual details and positions of small targets, aggravating the miss detection and false alarms. To address the above issue, we propose a Region-Guided Network with visual cues correction (RGNet) for IRSTD. More specifically, we design a Region Guidance Module embedded in shallow layers to generate the foreground mask by leveraging rich visual details contained in low-level features. The obtained mask then guides the re-weighting of deep feature maps to highlight the targets for further localization. Considering noisy signals in backgrounds tend to increase the false alarms of small targets, we propose a Visual Cues Correction Module, which extracts the regional features from low-level features by referring to the predicted positions of initial results, and conducts a binary classification to rule out the negative detection. Since the open-sourced IRSTD datasets are limited, we utilize both public and collected data for the evaluation. Both multi-target and single-target cases are investigated, and comprehensive experimental results indicate that compared to state-of-art models, our method achieves the overall best performance in both scenarios.