Research On Image Detection System Technology Of Communication Equipment Faults

Communication equipment is the infrastructure to ensure normal social and economic activities. Its production assembly process requires rigorous testing, which is usually conducted by experienced professionals. With human cost growing up, enhancing the efficiency of debugging and testing, lowering staff skill requirements and reducing commissioning costs are the major challenges faced by large equipment suppliers. Therefore, it’s urgent to develop a set of image diagnostic tools assisting people in equipment testing.Because equipment pictures are taken by intelligent terminal, position, angle and scale of target in the pictures will change. For fixed assembly test environment, the image registration and target segmentation methods have been studied based on feature points matching in this thesis. SIFT features of the selected standard image and the one under testing are calculated. These features of the interested region in the standard image should be matched with all features of testing image. According to the best four pairs of matching points, the perspective transform parameters are achieved. After registration, equipment area in test image coincides with which in standard image. At last, testing image can be segmented in accordance with positions of communication boards and Ethernet interfaces in standard image.In the test environment of variable assembly methods, positions of different communication boards are interchangeable. This thesis presents a detection method of tool boards based on HOG features, which is referenced to pedestrian detection. In order to obtaining the detectors, HOG features of various communication board samples are extracted as training data. In the form of sliding window, communication boards can be detected and the number of which can be counted, and then it is distinguished whether the assembly is correct or not. Meanwhile, because of large dimension of HOG feature vector, non-negative matrix factorization method is used for reducing HOG dimensions. This method can reduce SVM training time and tool-board detection time.The recognition of communication Ethernet interfaces faults will be used BP neural network and deep learning algorithms, the result of which have been compared. In BP neural network training, the gray edge and color edge of Ethernet interface picture are extracted. The edge images are divided into 25 parts and set the number of non-zero pixel value in each part as feature vectors. The results show that the performance of neural network classifier trained by color edge feature is better than gray edge feature. Two deep learning algorithms, convolution neural network and deep belief network, are used for learning Ethernet interface gray images to obtain interface state recognition classifiers. Deep learning algorithms performs better than BP neural network in classification.