Multi-Grained and Confidence-Aware Multiple Instance Network for Infrared Target Detection

Weakly supervised object detection (WSOD) methods that trains an object detection network using image-level labels has attracted much attention due to its cost-effective annotation and broad applied requirement. However, applying such weak label to detect targets in infrared images is not trivial due to the less discriminative target information and interference of complex backgrounds. This article proposes a multi-grained and confidence-aware multiple instance network (MCMIN) to detect infrared targets given the imprecise labels. The multiscale multi-grained feature extraction module is designed to capture discriminative features from different receptive fields for dim-small targets. The hierarchical multiple instance target detection module first applies L1-sparsity regularization to encourage the model generate reliable pseudo ground truth (GT), and then leverages the confidence-aware instance adaptive weighting strategy to refine proposals with particular emphasis, achieving more accurate target detection. The experimental results on two infrared target detection datasets illustrate that the proposed MCMIN outperforms other state-of-the-art WSOD methods with higher average precision (AP). The proposed approach decreases the false alarms.