Research On Human Visual Properties Introduced In Surface Defects Inspection And Classification Of Steel Strip

With developments in micro-electronics and computer technology, mimicking human-like behavior by machines has made great strides with good success as far as speed and reliability is concerned. Industrial machine vision systems to replace human inspection are growing trends in modern automation. The detection technology of machine vision with combination of sensors and active illumination realizes the function of getting image by human eyes. The image processing algorithm and intelligent recognition algorithm for data information processing replaces the human brain. High speed and high precision of detection requirement bring the huge data to vision system. How to improve system efficiency of large information processing and utilizing efficiency is the bottleneck that restricts development of vision inspection system. Surface defects of steel strip are characterized by varied types and complex patterns, which make real time and high reliability image processing algorithm a key subject of research in machine vision application.Study on the human visual understanding to bring the guidance and enlightening function on the machine vision research. The main research thrusts and achievements of this thesis are as follows:(1) The inspection method from the angle of image processing has fully developed, but the inspection research based on human vision system has just started internally. The inspection system of steel surface defects based on human visual properties simulates the human analysis and processing information to understand and analyze defect image. Expand machine vision system function to deal with complicated data, which has the function of promoted and improve to image processing algorithm.(2) The detection method based on gray information is not sensitive to the low contrast and small defects. So it’s difficult to judge the defects. The human visual system makes use of saliency mechanism to choose and reserve useful information from enormous input information. The saliency goal is be located quickly and the recognition is done accurately. A new detection method of steel strip surface defects is proposed. The detection model was built combining multiple scales and frequency of Gabor filter with multi-channel and multi-frequency of human visual system. The experiment results show that this method is effective to detect the surface defects and establish the coordinate position of defect region. The inspection speed is fast which can satisfy the real-time online detection system’s demand.(3) The two-dimensional grey images processed by vision system are the function of geometric qualities, illumination, surface characterizations, object color and camera parameters. So need research and analyze on vision system from different respects and different levels. Then the relevant resolve measures are presented. The double sensors are compatible with scattering and direct light to set-up hardware platform, which enhances image contrast and improve image resolution. Guarantee that the two-dimension and three-dimension defects are detected with no missing. By analyzing the direction of illumination influence on system performance, build the reflection mode of the texture surface and establish the relationship between height function and the change in the illumination angle. The experiment result shows that the error rate of segmentation method based on different direction illumination is less than the method based on grey statistical value.(4) The interference of noises, uneven light and complex background will bring great challenge to machine vision system. The illumination system is in the front of the visual inspection system, which affects the quality of captured image, complex degree of image processing and accuracy of classification. A new edge detection method based on order filters is presented to overcome the noise suppression and image details preserved. Relative to classical median filter algorithm, the anisotropic diffusion of TV image based on denoising method not only removes the noise and keeps details of image. In the qualitative and quantitative experimental analysis, PSNR and MSE show that the proposed approach is effective.(5) The images are recognized by human visual system through exacting image features to realize the expression of the abstract object information. The visual signal is decomposed by Gabor filter is similar to the way of identifying texture features by human visual system. So extract the image texture features through Gabor filter. The texture is the surface characteristics of object, which is not able to measure the essential attribute of object and can’t get the high level image contents only by using texture features. Combining Gray-level Co-occurrence Matrix extracts spatial information and neighborhood information to describe the scene. Imitate the two pathways property of human visual system, which will get better separation results of features space. Through adaptive weighting technology and vote technology, the recognition rate is improved efficiently. Test on five common types of defect, edge sawtooth, weld, inclusion, yellow spot, wrinkles and the results show that the feature dimensions are down70%and recognition rate is better than other common classifiers.