基于ResNet神经网络的激光光束系统质量检测的应用

Application of Quality Inspection in Laser Beam Systems Based on ResNet Neural Network

摘要: 激光技术凭借其高亮度、高方向性及高单色性等核心优势，已在现代社会诸多领域获得广泛应用。然而，当前光斑质量普遍存在良莠不齐的问题。在实际的科研生产时，由于光学元件像差、光束传输扰动等问题，导致光斑产生波前畸变与强度分布畸变，存在显著的高阶模式成分，光斑多偏离理想圆对称特征，呈现椭圆度或多瓣结构，影响了光束质量的准确评估。为突破上述技术瓶颈，本研究以深度学习技术为核心，选取预训练ResNet18模型构建迁移学习框架，以已构建的激光光斑样本数据集为训练样本库，通过特征迁移与模型应用，对激光光斑质量进行评估。实验验证结果显示，该模型对激光光斑质量评估的准确率达80%，可有效识别光斑质量缺陷，显著提升光斑的光强均匀性与形态稳定性。

Abstract: Leveraging its core advantages of high brightness, high directionality, and high monochromaticity, laser technology has been widely adopted across various fields in modern society. However, the quality of laser spots currently exhibits significant inconsistency. In practical scientific research and production processes, factors such as optical aberrations and beam transmission disturbances lead to wavefront distortions and non-uniform intensity distributions in laser spots. These often manifest as prominent higher-order mode components, deviations from ideal circular symmetry, and the emergence of ellipticity or multi-lobe structures, thereby impeding accurate assessment of beam quality. To address these technical challenges, this study employs deep learning as the core methodology, utilizing a pre-trained ResNet18 model to construct a transfer learning framework. A self-built dataset of laser spot samples serves as the training database. Through feature transfer and model application, the proposed approach evaluates laser spot quality. Experimental results demonstrate that the model achieves an accuracy rate of 80% in laser spot quality assessment, effectively identifying spot quality defects while significantly improving intensity uniformity and morphological stability.