视觉焊缝识别因其精确、快速、可靠及可数字化等优点，成为了近年来焊缝识别领域的热点研究问题，在焊接工艺、焊缝缺陷检测、焊缝智能识别等方面有着巨大的应用前景。在视觉焊缝识别定位中，环境光照的变化与腐蚀磨损是阻碍识别性能提升的主要因素，因此，解决光照及腐蚀磨损问题成为了焊缝识别定位的关键。 本文主要对焊缝识别定位中的特征提取、分类训练、识别定位三个关键技术进行研究。在特征提取的研究中，采用基于图像块的选取方式对焊缝图像特征进行提取，并利用PCA对特征进行降维，利用累计贡献率曲线为降维参考，以曲线放缓点维度作为各尺寸图像块特征的确定维度，降低了图像块特征维度。在分类训练的研究中，采用基于LM的BP神经网络作为分类器，利用公式法和增长法结合的改进方法对网络隐层节点进行确定，优化了隐层节点的选择。识别定位依据分类结果对图像进行重构，利用横向灰度值平均、最小焊缝宽度为识别定位指标，确定焊缝中心线。 本文在Matlab上搭建了一个基于图像块特征的焊缝识别系统，并在不同成像距离、不同光照的条件下对识别系统的性能进行了评估，实验结果表明：该系统对于在不同光照条件下，具有不同程度腐蚀磨损因素时，有重叠识别的检测率为90.5%，虚警率为7.4%，无重叠识别的检测率为85.9%，虚警率为4.7%。

With the advantages of accurate, fast, reliable, easy to digitize, etc., visual welding seam recognition has been the research hotspot in the fields of weld recognition in recent years, and has great application prospect in the aspects of welding process, weld defects detection, welding seam identification, etc. In visual welding seam recognition, the influence of illumination change and the corrosion are the key factors that hinder the recognition performance. In this paper, the researching and improvement of the algorithms is conduct around three aspects: feature extraction, training of the classification, welding seam recognition and positioning. In the part of feature extraction, we extract features based on image block, and use PCA to reduce the dimension of image block features. Using the cumulative contribution rate curve as a dimension reduction reference, which choose the curve slow point as a choice for each image block features. In classification training section, we choose the BP neural network based on LM as a classifier. In order to optimize the choice of the hidden layer nodes, we use The Comprehensive Method based on formula method and growth method to determine the network hidden layer nodes. For the part of welding seam recognition and positioning, according to the classification results, we reconstruct recognition image, using the horizontal gray value average and the minimum weld width as a positioning index, determine the weld centerline. On the basis of the above study, we build a welding seam recognition and positioning system on MATLAB, which can process welding seam images. At last we conduct experiments to evaluate the performance of the system in three different imaging distance and different light conditions. The experimental results prove that the overlapping detection rate is 90.5%, the false alarm rate is 7.4%, the non-overlapping detection rate is 85.9%, the false alarm rate is 4.7%.