COD-YOLO: An Efficient YOLO-Based Detector for Laser Chip Catastrophic Optical Damage Defect Detection

Abstract

High-power semiconductor lasers play a crucial role in optical communication systems, and their reliability is key to the normal operation of the system. Catastrophic Optical Damage generated during operation is a major factor affecting chip performance and lifetime. Accurate detection of the location and development process of damage, along with the study of failure mechanisms and degradation modes, is a pressing issue. We propose an intelligent analysis approach based on the YOLO architecture for defect detection in laser chip Catastrophic Optical Damage, named COD-YOLO. To overcome challenges such as the similarity of defect features, complexity of background features, and inaccurate spatial positioning, the network employs deformable convolutional and channel attention. This adaptive approach captures rich feature representations and simultaneously addresses long-distance dependencies and adaptive spatial aggregation. Combining spatial-content-based upsampling in model neck achieves multiscale feature fusion, improving perception and understanding through the integration of semantic and positional information. Furthermore, due to the lack of fine-grained information, IoU metrics are highly sensitive to the positional deviation of tiny defects. Combining tiny object detection loss function to measure the regression of bounding boxes, adapting to variations in defect scales, experimental findings demonstrate that COD-YOLO outperforms other competing methods in detecting Catastrophic Optical Damage in the active region of the laser chip.