Aiming at the problem of low detection accuracy in traditional semiconductor laser chip defect detection, a semiconductor laser chip defect detection and classification method based on improved Faster R-CNN is proposed. Firstly, a semiconductor laser chip defect acquisition device was built to collect the state of the chip during operation and establish a dataset; then the ResNet50 feature extraction network was optimised to reduce the residual block structure while using multiple 3*3 convolutional layers to improve its ability to detect important information; finally, the CA (Coordinate Attention) attention mechanism was introduced into different layers of the improved Res-Net50 network in different layers to adaptively learn the importance weights of each channel and further improve the feature representation ability. The experimental results show that compared with the original network, the proposed method improves the detection precision and classification accuracy, has better recall and accuracy, and is able to quickly and accurately carry out defect detection to further improve the process.