Blurred and occluded target recognition in terahertz images based on improved YOLOv5

Terahertz (THz) imaging technology is widely used in applications such as security screening, radar detection, and biomedical applications. Nevertheless, due to the inherent limitations of imaging conditions, THz images often exhibit low contrast, blurred contours, and reduced feature information under partial occlusion, which significantly compromises recognition accuracy. To address these challenges, we proposed an Enhanced and Occlusion-aware Focus YOLOv5(EOF-YOLOv5), an improved architecture based on YOLOv5. In this study, image enhancement preprocessing was applied to raw THz image datasets acquired from a THz active array imaging system to improve target contrast and contour clarity. An Occlusion Aware Context Attention (OCA) mechanism was integrated into the neck network of YOLOv5. This mechanism dynamically adjusts attention weights and enhances feature responses in visible areas by capturing spatial occlusion patterns via 1 × 1 convolution and modeling inter-channel dependencies. Additionally, the original Complete Intersection over Union (CIoU) loss function was replaced with the Focal-Efficient Intersection over Union (Focal-EIoU) loss function to reduce excessive focus on simple samples and improve the learning performance for challenging samples. Experimental results demonstrate that image enhancement preprocessing significantly improves both visual quality and structural information, thereby boosting the network's recognition accuracy. On the same preprocessed dataset, the EOF-YOLOv5 algorithm outperforms the baseline YOLOv5 model, elevating precision (P) from 66.7 to 79.3%, recall (R) from 73.1 to 80.6%, and mean average precision (mAP50) from 75.5 to 83.7%. The proposed model effectively identifies targets under blurry and occluded conditions, providing an innovative solution for terahertz image detection.