关键词: 绝缘子缺陷, 多缺陷检测, 多维协作注意力, 小目标

Abstract: Aiming at the problems of small detection targets, large scale differences, multiple types, and complex backgrounds in aerial insulator defect images of transmission lines, a cross-scale multi-dimensional collaborative feature interaction algorithm for aerial insulator multi-defect detection is proposed. In the process of multi-scale feature fusion, a multi-dimensional collaborative attention module is introduced to enhance the feature interaction between spatial and channel dimensions, thereby enhancing the model??s global perception ability towards small targets. Based on this, the C2f-CM module is proposed, which effectively reduces feature structure redundancy and improves the model??s ability to extract cross-scale features through the segmentation and transformation of feature channels and the split fusion of extended branches. A global receptive field-space pooling pyramid fast module is constructed to help the network suppress interference from complex backgrounds and aggregate insulator defect features from different receptive fields. By designing the MG-Detect detection head, the missed detection of small targets is reduced, the algorithm's adaptability to multi-scale targets is enhanced, and redundant calculations are reduced. The experimental results show that the algorithm??s mAP@0.5 reaches 97.3%, an increase of 5.4?percentage points compared to the baseline model. Meanwhile, the AP for insulator break, flashover, and self_blast defects are increased by 5.1, 9.9 and 1.1?percentage points, respectively. This algorithm can more accurately identify the characteristics of insulator defects at different scales.

Key words: insulator defects, multi-defect detection, multi-dimensional collaborative attention, small goals