Online inspection of narrow overlap weld quality using two-stage convolution neural network image recognition

摘要

In narrow overlap welding, serious defects in the weld will lead to band breakage accident, and the whole hot dip galvanizing unit will be shut down. Laser vision inspection hardware is used to collect real-time image of weld surface, and image defect recognition and evaluation system is developed to real-time detect quality. Firstly, region division is implemented. In view of the characteristics of gray image such as large information, low contrast and blurred edge, an improved image segmentation algorithm is proposed by using image convolution to enhance edge features and combining with integral image, which can quickly and accurately extract the weld edge and divide the region, and the processing time can meet the real-time requirements. Then comparing the effect of traditional method and convolution neural network in identifying defects, VGG16 is used to identify weld defects. In order to ensure real-time performance, a two-stage weld defect recognition is proposed. First, the large defective area is identified, and then, the defect is accurately identified in the area. This method can quickly extract defect areas and complete weld defect classification. Experiments show that the accuracy can reach 97% and average running time takes 3.2 s, meeting the online detection requirements.