Multi-scale self attention based method for automatic defect detection in electroluminescent images of semiconductor laser chips

As the pivotal element in optical transmission systems, the semiconductor laser holds paramount significance in enabling the next-generation of high-speed optical communication, characterized by cost-effectiveness, minimal power consumption, and unparalleled capacity. The occurrence of laser failure stemming from chip damage poses a significant constraint on coherent transmission performance, thereby necessitating a deeper exploration into the underlying damage mechanisms and the development of sophisticated damage detection methodologies. This endeavor is crucial not only to preemptively prevent damage but also to streamline the process of subsequent repair, ensuring the seamless operation and advancement of optical communication technologies. Efficient recognition and precise localization of catastrophic optical damage (COD) defects are crucial steps in advancing laser chip manufacturing, significantly enhancing chip lifetime and output power. Traditional manual methods are often subjective, labor-intensive, and heavily reliant on human judgment, which can be time-consuming. Machine learning-based approaches have shown promise in COD defect detection, demonstrating superior performance compared to manual methods. However, these methods still face challenges in accurately detecting tiny, irregular, and weak texture defects. To address the aforementioned issues, in this paper, we proposed an innovative damage detection network and built a COD segmentation dataset based on electroluminescence images. A feature aggregation and enhancement module (FAEM) based on convolutions of various sizes is introduced before the encoder to capture diverse features effectively. A contextual semantic features acquisition module (CSFAM) consisting of dilated convolutions with multiscale expansion rates and pooling layers with various sizes of pooling kernels is introduced between the encoder end decoder to encode global contextual information while preserving spatial details. Furthermore, self-attention modules are also introduced to capture the channel and spatial relationships, facilitating better fusion of the COD features across different scales and resolutions. The experimental results show that the proposed network in this paper achieves an accuracy value of 99.881, a Dice score of 85.838, and a Re value of 88.653 on the COD dataset. These metrics surpass those of sophisticated segmentation methods such as DenseASPP and DeeplabV3, highlighting the robust performance of our approach in detecting and segmenting COD defects accurately.