Research On Surface Defect Recognition And Detection Technique Based On Machine Vision And Artificial Neural Network

1. Purpose and significance of the issueIn industrial production, in order to eliminate shoddy products and identify the fault-related hidden troubles, we need to check the quality of products. Ningbo is China's bearing production base. There are a small amount of bearings which are detected as rejects and shoddy products in the finished products due to the pits, scratches and some other defects on their outer surfaces. These substandard products must be removed before delivery. At present, enterprises adopt the artificial visual method to inspect produced bearings one by one. The work is not only heavy, low efficiency, but also has a high missing rate in the inspections. The research of this issue is of importance significance and has a high practical value to resolve the problems and realize bearing automatic detection. Because the size, depth and location of the pits and scratches on the surfaces of micro bearings are random, it's the best way to solve this problem by using CCD image recognition technology to make non-contact detection. In this paper, CCD is used to acquire images of bearing surfaces and digital image processing technique is used to carry out the pre-treatments of the bearing images to separate the targets for detection. Then make feature extraction of the images after pre-treatments and use artificial neural network method to judge whether the bearings are eligible or not. Results of the judgments can be regarded as the bases to remove the unqualified part in the next step. By automatically removing unqualified parts and retaining qualified parts according to the inspection results, the system achieves automatic detection of the defects on micro bearings'surfaces.2. Overall program of detection systemThe micro bearing detection system of this paper is based on the machine vision inspection system. The overall program plan of the detection system is shown in Fig.1. The hardware of the detection system in this are: light source, lens, CCD camera, image acquisition card and visual processor. By considering the surface structure, size, detection accuracy, characteristics of defects, and actual experimental conditions, the composition of micro bearing detection system hardware, as shown in Table 1,can be determined according to the principle of hardware selection. Table 1 Hardware composition of detection system light source lLigEhDt (sLoDurRce2 -o9f0 lSoWw-)a ngle annular CCD A102f-type Lens TEC-M55 telecentric lens image acquisition card 1394 digital image acquisition card computers tthhee shloasvte c coommppuuteter:r :E PVLOCC o-ft yFpPe0 i-nCd1u4s trial computer3. Image preprocessingImage preprocessing including image filtering, image thresholding segmentation, image edge detection and target image extraction. In this paper, median filter is used to remove the noise in images; the method of iterative threshold is carried out to separate the end of sealing cover and complete image binarization; Roberts operator is used for edge detection; Hough transform Principle is used to find the center position of micro bearing images and then separate detection targets from the original image. Image preprocessing process is shown in Fig.2. 4. Image recognition and the establishment of detection modelThis paper adopts artificial neural network in pattern recognition and makes combined moment of image as neural network input. Maitra has given a method of using six invariants based on seven moment ones with translation, size and rotation invariance to achieve target classification. This paper chooses five combined moment invariants composed of these six to realize image feature extraction and set them as the neural network input. Then calculate combined moment invariants of multiple images to establish image feature library.This paper aims to establish BP neural network detection model. Because a continuous function within any closed interval can approximate by a BP network with one hidden layer, this paper selects a BP neural network containing one hidden layer. In the micro bearing detection model of this paper, the eigenvectors are composed of five images combined moment invariants. Test result of the model is just to judge whether the parts are eligibility. According to the number of hidden layer by empirical formula as well as the actual experiment, the paper determines model structural diagram of BP neural network and the parameters as shown in Fig.3 and Table 2. Fig.3. Model structural diagram of BP neural network Table 2 Parameters of BP neural network Network layers Input layer nodes Hidden layer nodes Output layer nodes Transfer function 3 5 6 1 OHuidtdpeunt llaayyeerr:: ptaunresligin This paper uses trainlm as training function of the network model. The specific training parameters are shown in Table 3. Table 3 training parameters The largest number of training 2000 Precision of training 10-7 The largest number of failures 5 Minimum gradient requirements 10-6 Showing iterative process of training 25 Initial value of m 0.001 Reduce rate of m 0.1 Growth value of m 10 Maximum value of m 1010 After 147 times training, the network achieved the scheduled approximate accuracy.5.Verification of the model testing capabilitiesChoose 25 images including 15 qualified and 10 failed from the image library as the test samples of the model. After testing, there were two of the 15 original qualified images were misidentified as unqualified while the 10 original unqualified images were all identified as unqualified as shown in Table 4. As shown in Table 4, the correct identification rate of the model is 92 percent, which fully shows that the neural network detection model of the micro bearing surface defects has high accuracy rate, reliability and good safety.