The Based On Improved Image Saliency Characteristic Detection Billet Surface Defect Detecting Technology

With the continuous development of computer vision technology, industry-orientedvisual defect detection technology has attracted more and more extensive attentionincreasingly. However, as the production technology’s rapid ascension, many traditionaltechnologies have been unable to meet the requirements of industrial production. If thedefect detection technology can simulate the way of vision system working, whichprioritized locating suspected defect areas and then centralized processing which can avoidcomputing resources waste and improve detection accuracy. Therefore, how to design thevisual defect detection process that conforms to the saliency features of visual attention hasan important practical value.First of all, this thesis elaborates the composition, attention mechanism and formingprocess of the human visual system. Then the thesis introduces the existing classicalsaliency model in detail, including the based on biological vision Itti model, the based onfrequency domain fast processing Hou model, the based on context-aware Gofermanmodel and the based on image segmentation and area contrast Cheng model. These models,however, also have limitations: low efficiency, feature information simplification and poorrobustness to complex background and so on.Synthesized the requirements of visual defect detection, this thesis presents a novelsaliency algorithm based on the fusion of global chromatic difference and local low-levelfeature. Simulated the visual system mechanism, the method calculates the color saliencyin nature scene image to get the global saliency feature map at first, which combines thehistogram quantization and saliency smoothing; then extract a variety of low-level features,make them fusion on multi-scale for different features and get final local saliency map inlinear weighted synthesis way; Finally through the fusion of global and local features and the strategy model of feedback to the global feature restrain background interference, wecould get final saliency map. This saliency algorithm could position saliency target morequickly and accurately and effectively restrain the influence of complex background,which is complied with the actual application environment of visual defects detection. Wetest this algorithm in Achanta public test data set and compare with some classical saliencyalgorithm. The result, which includes computing speed, visual effects, ROC curve and areaunder curve, precision-recall-f measure value, shows that our saliency algorithm is at ahigher level, which basically meets the requirements of defect detection such as highprecision, high speed and so on.At last, this thesis designed a new billet surface defect detection method based on thesaliency features. Combined the proposed saliency features with the edge features fromGabor wavelet transform, we could get the high credibility suspected defect area, therebybeing conducive to reduce Adaboost detection and positioning time. The experiment resultsshow, the new method of reliability is high, algorithm efficiency is improved obviously,and the method has good practical application value.