Enhancing Power Line Insulator Segmentation in Aerial Imagery with Integrating YOLOV5 and SAM Zero-Shot Learning Method

This paper presents an innovative approach to advance defect detection and segmentation in power line insulators integrating the YOLOV5 detection framework with the SAM zero-shot segmentation method. The conventional methods of defect detection, often limited by predefined defect types, lack the adaptability to newly emerging defects. Leveraging the advanced zero-shot learning's capabilities, our novel integration overcomes this challenge, successfully identifying and segmenting defects not included in the training data. Employing a robust data pipeline supported by the high-quality masks and annotations of the Comprehensive Power Line Insulator Dataset (CPLID), we demonstrate the versatility and efficiency of our approach. An extensive and rigorous evaluation using the 5-fold cross-validation and Dice Similarity Coefficients affirmed our model's heightened overlaps accuracy in segmentation tasks. The presented solution promises remarkable potential for adoption and adaptation across various contexts beyond power line insulators, hence setting a significant milestone in aerial imagery analysis and applications in computer vision.