Catastrophic optical damage (COD) is one of the crucial factors severely constraining the performance of high-power lasers. An accurate COD defect location is of great significance to laser chip manufacturing, which could be used to improve the production process and optimize the structural design of laser chips. A manual detection method for laser chips is very time-consuming and costly. Recently, deep-learning-based methods have demonstrated outstanding performance in various fields, owing to their robust feature extraction capabilities. However, these methods still have limitations on the samples with weak texture, class imbalance issues, and random size of targets. To address these issues, a novel COD defect segmentation method is proposed. And electroluminescence imaging technology is utilized to visualize defects inside the laser chip and collect the COD dataset. To improve the extraction capacity of the strip-like COD feature, a rectangular dilated convolution is proposed to increase the receptive field of the convolution. To acquire richer information from local contextual features, a multiscale feature aggregation block (MFAB) consisting of multiscale rectangular dilated convolutions is introduced to acquire multiscale feature maps. An attention module is applied in the proposed block to highlight the defect features. Moreover, to enhance the segmentation capacity on random-scale defects and class imbalance issues, a deeply supervised side-outputs fusion block is proposed to fuse multiple side outputs at different semantic levels to generate the final segmentation map, which is used to improve COD detection performance in the way of feature pyramid. Experimental results on the COD segmentation dataset demonstrate that the proposed method outperforms other state-of-the-art segmentation methods.

中文翻译：

---

MFSF-Net：用于像素级灾难性光学损伤检测的多尺度特征和侧面输出融合网络

灾难性光学损伤（COD）是严重制约高功率激光器性能的关键因素之一。准确的COD缺陷定位对于激光芯片制造具有重要意义，可用于改进生产工艺和优化激光芯片的结构设计。激光芯片的手动检测方法非常耗时且成本高昂。最近，基于深度学习的方法由于其强大的特征提取能力而在各个领域表现出了出色的性能。然而，这些方法对于纹理较弱、类别不平衡问题和目标随机大小的样本仍然存在局限性。为了解决这些问题，提出了一种新的 COD 缺陷分割方法。利用电致发光成像技术可视化激光芯片内部的缺陷并收集 COD 数据集。为了提高条状COD特征的提取能力，提出了矩形扩张卷积来增加卷积的感受野。为了从局部上下文特征中获取更丰富的信息，引入了由多尺度矩形扩张卷积组成的多尺度特征聚合块（MFAB）来获取多尺度特征图。在所提出的块中应用注意模块来突出缺陷特征。此外，为了增强对随机尺度缺陷和类别不平衡问题的分割能力，提出了一种深度监督的侧输出融合块，以融合不同语义级别的多个侧输出以生成最终的分割图，用于改进COD检测以特征金字塔的方式表现。 COD 分割数据集上的实验结果表明，所提出的方法优于其他最先进的分割方法。