基于改进ResNet34卷积神经网络的硅棒缺陷检测
2025,33(2):37-43
摘要：针对现有人工检测太阳能硅棒导致准确率低、识别速度慢的问题,提出一种基于改进深度学习网络的硅棒外观缺陷分类系统；该系统把硅棒检测分为两个部分,运用图像算法对硅棒图像进行预处理,结合阈值分割和连通域分割方法对硅棒图像区域分割,根据区域轮廓采用内接圆的方法将硅棒从图像中提取,减少特征提取干扰因素；其次改进残差网络的残差模块的结构,提高网络训练速度；实验模型加入正则化和Dropout层改善过拟合现象,采用改进余弦退火的学习率衰减方式寻找模型的最优解；实验结果显示,与未改进的残差网络相比,改进的网络总体识别准确度提高2.41%,该模型有效的提高硅棒外观缺陷分类的高效性和泛化能力。
关键词：外观缺陷;阈值分割;残差网络;L2正则化;余弦退火;
Small Sample Acoustic Target Recognition Method Based on Deep Learning
Abstract：Aiming at the problems of low accuracy and slow recognition caused by manual detection of solar silicon rods, a silicon rod appearance defect classification system based on improved deep learning network was proposed. In this system, the silicon bar detection is divided into two parts. The image algorithm is used to preprocess the silicon rod image, and the threshold segmentation and connected domain segmentation are combined to segment the silicon bar image region. The silicon bar is extracted from the image according to the region contour by the method of inset circle to reduce the interference factors of feature extraction. Secondly, the structure of residual module of the residual network is improved to improve the training speed of the network. regularization layer and Dropout layer are added to the experimental model to improve the overfitting phenomenon, and the learning rate attenuation method with improved cosine annealing is used to find the optimal solution of the model. The experimental results show that compared with the unimproved residual network, the overall recognition accuracy of the improved network is increased by 2.41%, and the model can effectively improve the efficiency and generalization ability of silicon rod appearance defect classification.
Key words：appearance defects; threshold segmentation; Residual network; regularization; cosine annealing;
收稿日期：2023-12-11
基金项目：
