Research Of Pattern Recognition Based On BP Neural Network For Vision-Inspecting System

Industry vision-inspecting system can be divided into three classes: inspecting, orientation and assemblage. This system belongs to the class of inspecting, in other words, it can classify and recognize the industrial components, and distinguish whether the part is good up to grade or not. In industrial production, the determinacy of the industrial environment and illumination made the idea possible that the industrial components can be distinguished and classified rapidly and accurately. As a result of the idea this system is developed. System adopted CG card and image sensor on gathering the original image, which avoids contacting with objects directly and manual inspecting; the system made use of VC programming the part of image processing, recognition algorithm adopts ANN, which can simulate many kinds non-linearity mapping on the bases of the shape of industrial parts. It's important and difficult of abstracting fit characters in recognition system. For a map of digital image, which is recognized, we must do some processing and analysis to it if we want to abstract the characters efficiently. In recognition the characters that are often elected are image range character, image statistics character, image geometric character, image transform coefficient character and so on. Among these character, histogram character, statistics character, area, perimeter, form factor and eccentricity etc are abstracted in this task. These characters, which are compressed, would be as sample training ANN. For a trained ANN if we input the characters it can output the result of recognition, accordingly achieved the aim of recognition. In all kinds of model of ANN, the multi-layer feed-forward network was used most successfully and most excessively. Further more, the BP network is most representative. Due to the network adopt supervised training method, so it can only used in supervised recognition. Although BP neural network acquired a certain extent succession in practice, it existed many limitations. Such as: existing partial minimum, the slow convergence, selecting the number of hidden nodes difficult and so on. Aim to the specialty of traditional BP network, For improving the speed and practicability of ANN in recognition, in this paper, a back-propagation neural network model has been proposed to produce a more effective recognition based on adjusting the connection mode of the network. In addition to, an algorithm and feature extraction corresponding to the improved ANN also has been proposed. Experiment results show that the weights initialization becomes easier, convergence and recognition speed are increased; the improved network reduced to fell into the possibility of partial minimum effectively; the network adopts multi-output coding for recognition, has some adaptability.