Semantic segmentation of mechanical parts based on fully convolutional network

Machine vision detection and recognition technology has a wide range of advantages such as non-contact, real-time online, fast, strong anti-interference ability. This technology has been used to achieve zero inferior parts production of mechanical parts, to meet the requirements of modern manufacturing progress and development. In the actual production it shows a broad application prospects. In this paper, we apply the Fully Convolutional Networks (FCN) method to the robot vision system for mechanical part inspection. We found that the use of FCN can make the classification accuracy quickly reach a stable value, thus greatly reducing the training time. The proposed method is based on (1)our own database is collected on an industrial conveyor belt, which contains 2248 images of various types of parts with different sizes, (2) Fully Convolutional Networks(FCN) is trained by using back-propagation algorithm to extract valuable features from input training data, and realized pixel-level semantic classification on the validation images.