A Terahertz Image Target Recognition Method Based on Improved YOLOv5

Terahertz (THz) waves exhibit a number of advantageous properties, including low radia-tion, and penetration capability, in addition to a broad bandwidth. These characteristics render terahertz image target recognition a prominent area of research. To address the is-sue of insufficient recognition accuracy of terahertz images when targets are blurred and features are limited, we introduce Enhanced and Occlusion-aware Focus YOLOv5(EOF-YOLOv5), an improved model based on YOLOv5. This study performs im-age enhancement preprocessing on raw terahertz image datasets acquired by a terahertz near-field array imaging system to enhance target contrast and contour information. For the improved network EOF-YOLOv5, first, an Occlusion-Aware Context Attention (OCA) mechanism is integrated into the neck network of YOLOv5, which employs deformable convolution to predict occluded regions, dynamically adjusts attention weights, and en-hances feature responses in visible areas. Second, the original Complete Intersection over Union (CIoU) loss function was replaced with the Focal-Efficient Intersection over Union (Focal-EIoU) loss function to mitigate excessive focus on simple samples and enhance the learning capability for challenging samples. The experimental results demonstrate that image enhancement preprocessing enhances both visual quality and structural infor-mation while improving network recognition accuracy. On the same preprocessed dataset, the EOF-YOLOv5 algorithm achieved a 12.6% increase in precision (P) to 79.3%, a 7.5% improvement in recall (R) to 80.6%, an 8.2% boost in mean average precision (mAP50) to 83.7% compared to the baseline YOLOv5 model. Compared with other mainstream algo-rithms, the proposed algorithm achieves more accurate detection of object locations, vali-dating its feasibility and effectiveness.