Dynamic Context-Aware Pyramid Network for Infrared Small Target Detection

Detecting faint and diminutive infrared targets devoid of clearly defined shape and texture details in intricate surroundings remains a formidable challenge within the domain of target detection. Current methodologies employing deep neural networks and pooling operations can easily cause small target loss, resulting in suboptimal detection outcomes. To address these issues, we design an innovative dynamic context-aware pyramid network (DCAPNet). It comprises three core modules: dynamic context modulation (DCM), dynamic pyramid context (DPC), and shuffle attention fusion (SAF). Specifically, the DCM module is designed to adaptively capture diverse-scale information from input images, adapting to various target dimensions and enhancing the feature representation capabilities crucial for effective target detection. Subsequently, the DPC module adaptively captures multi-scale features and better utilizes contextual information by aggregating multiple DCM modules. This facilitates the retention of essential semantic information about small infrared targets within deeper network layers. Finally, through the designed SAF module, we facilitate the exchange of information within the same layer and establish correlations between different layers, ensuring the fusion of shallow spatial positional information and deep semantic information to enhance the overall detection performance. Furthermore, comprehensive ablation studies are conducted to substantiate the efficacy of the designed modules within the proposed network architecture. Simultaneously, we conducted a comparative analysis of the proposed network algorithm against several state-of-the-art methodologies for infrared small target detection, employing multiple evaluation metrics. The results consistently demonstrated the proposed model attains superior detection performance on the publicly available IRSTD-1k, SIRST-Aug and NUDT-SIRST datasets.