Abstract: Aiming at the problems of insufficient feature perception, poor multi-scale feature fusion, and limited bounding box localization accuracy in infrared small target detection, this paper proposes an infrared target detection method based on micro-scale extraction and multi-scale fusion (CAANet). Based on YOLOv8n as the baseline model, a Coordinate-Attention Enhanced Micro-Scale Feature Extraction Module (CMS Block) is designed to enhance tiny target feature perception through multi-scale depthwise separable convolutions and a coordinate attention mechanism. An Adaptive Spatial Feature Fusion module (ASFF) is introduced to dynamically schedule multi-scale information and improve fusion consistency. An Adaptive Infrared Intersection over Union (AIoU) loss function is proposed to optimize localization accuracy by integrating spatial geometric constraints with grayscale saliency features. Experimental results on the IRSTD-1k dataset demonstrate that CAANet achieves a precision of 80.2%, a recall of 81.1%, and an mAP@50 of 83.8%, significantly outperforming mainstream detection models. The robustness evaluation on the NUDTSIRST dataset further validates its generalization capability, indicating that the proposed method can provide effective technical support for infrared small target detection tasks.