On-line Surface Defect Detection Of High-temperature Slabs Based On Multi-information Fusion

Surface defects of continuous casting slabs, such as cracks, are not only the main factors of affecting the production process of continuous casting, but also the main reasons of causing quality accident of hot-rolling. On-line detection of surface defects of high-temperature slabs can avoid successive quality accidents, and provide technical support for the process of Continuous Casting-Direct Rolling (CC-DR), which can greatly reduce the energy consumption of steel rolling. Most of off-the-shelf on-line surface detection systems use two-dimensional detection, which is realized by gray-scale images recognition of surface. However, there are a lot of scales and slags on slabs, which are very similar to real defects in gray-scale images, and it is very hard to distinguish them with two-dimensional detection. Therefore, false-detection is main problem of two-dimensional detection in surface inspection of slabs.Feature extraction methods of surface images for continuous casting slab were studied, and a new method of three-dimensional detection of surface defects based on photometric stereo was developed. Surface defects of slabs were detected and recognized through fusion of gray-scale information acquired by traditional ways and depth information acquired by photometric stereo. The main results and innovations of the dissertation are as follows:(1) Methods of multi-scale geometric analysis were applied to feature extraction of surface images for slabs. Different types of defects have special information at some directions and scales of their images, while the traditional methods of feature extraction, such as Wavelet transform, are unable to get the information at all directions. Curvelet transform, Contourlet transform and Shearlet transform were applied to multi-scale and multi-direction decomposition of surface images, statistical features, such as means and variances of the sub-band images acquired by decomposition were calculated to compose a high-dimensional feature vector. Supervised Locally Linear Embedding (SLLE) was applied to remove redundant information of the high-dimensional feature vector, which was transformed to a low-dimensional feature vector. The low-dimension feature vector was fed into a classifier based on support vector machine for defect classification. The methods were experimented with samples from production lines of continuous casting, and the recognition rate achieved by Shearlet-SLLE based on Shearlet transform and SLLE is up to87.36%, which was higher than that achieved by Wavelet transform, Curvelet transform and Contourlet transform.(2) A three-dimensional detection scheme for slabs was designed based on color photometric stereo. A reflection modal of slabs was measured with camera calibration Athree-dimension detection scheme was designed based on color photometric stereo, which was realized by combination of red, green, blue laser line sources and one line-scan camera with three CCD arrays. A new algorithm of three-dimension reconstruction based on quad tree integral was developed, and depth values were calculated from gradient fields. Compared with the traditional algorithms of rectangular path integral and global optimization, the algorithm had higher efficiency, and higher robustness to noises. Results of experiments showed that the algorithm met the time requirement of on-line detection, and had low errors of reconstruction.(3) Detection and recognition of surface defects based on the information fusion of gray scales and depth values was proposed. According to datavolume of gray scales, gradients and depth information, a three-layer algorithm of information fusion was proposed. The algorithm achieved optimization in efficiency and accuracy. An experimental system of surface defect detection for continuous casting slabs was developed based on information fusion of gray scales and depth values. The system consists of on-line recognition function and off-line training function. The on-line recognition function provided image samples for off-line training, and the off-line training function trained the classifier of on-line recognition. The on-line recognition function used information fusion algorithm of neural network and fuzzy sets. The off-line training function trained four independent classifiers respectively, whose parameters were used in the on-line recognition function. The system was tested with430samples from a production line of continuous casting, including defects cracks, melting lines, scales, and so on. The results showed that the overall defect recognition rate was up to95%, and the average operation time for one image of4096*512is2.23s, which met the requirement of on-line detection.