Research on Machine Vision Defect Detection Algorithm Based on Deep Learning

This paper conducts a detailed exploration of defect recognition algorithms in machine vision supported by deep learning, mainly focusing on the performance and potential optimization paths of Retina Net in industrial environments. This paper conducts a comparative analysis of various convolutional neural network architectures to improve the recognition accuracy and adaptability of the model. At the same time, customized algorithm tuning is implemented based on the types of defects that frequently occur in industrial production, such as surface scratches, cracks, and unevenness. The industrial site's complex background and multidimensional target characteristics prompted this paper to innovate the multi-level feature pyramid structure of Retina Net, significantly enhancing its performance in the recognition of small objects and complex patterns. In the simulation test phase, the effect was verified with the help of actual industrial data sets. The data showed that the optimized and upgraded Retina Net algorithm surpassed traditional technologies in multiple evaluation indicators, especially in recognition accuracy, with an increase of about 7.8%, while the false alarm rate was reduced by 12.3%. In addition, the algorithm showed excellent performance in the recognition task of multi-scale targets, with an average recognition time reduction of about 15%. The experimental results powerfully demonstrate the effectiveness and excellent performance of the improved algorithm in real-world applications.