Research Of Surface Detection And Classification Based On Machine Learning

Target detection and recognition is one of the important research fields of machine vision.From the use of traditional image processing methods,the combination of machine learning on the analysis of the target,and then to deep learning network based applications in recent years,the situation is getting better and better in this field.By 2019,some related products have appeared in people's daily work and life,not to mention its application in military and industrial.In simple terms,target detection and recognition is to obtain the location and category information of interested targets by classifying the information contained in the image itself,so as to achieve the purpose of obtaining relevant information in the scene of intelligent machine analysis.The object detection and recognition studied in this paper is mainly aimed at some small parts in the industry.The purpose is to accurately detect and identify the object from the complex background image,laying a foundation for the subsequent automated auxiliary testing.The algorithm proposed in this paper mainly uses the model generated by the combination of target features and classifier training to detect and recognize the processed images.The main contents of this thesis are as follows:1.The general principle of target feature and extraction method is studied,and an improved method based on LBP and maximum deviation criterion is proposed for the characteristics of industrial parts.After testing,the method combined with classification training to get more accurate target candidate region calibration results,less error calibration region,the improved algorithm to describe the image texture features more suitable for the target in this thesis.2.This paper studies the image edge sub-pixel detection based on the feature of Zernike moment.Aiming at the problem of inaccurate description of general edge features,this paper proposes a feature description method based on the template of Zernike 7×7.After testing,the target edge features extracted by this method are more refined,reaching sub-pixel level,the edge information obtained by statistics is more abundant,the detection effect after combining with the classifier training is better.3.The fusion method of texture feature and edge feature was studied.Aiming at the problem that a single feature could not describe the target feature more comprehensively,a fusion method based on improved LBP feature and Zernike moment feature was proposed.Experimental results show that this method is better than the previous single feature detection method in error region discrimination.4.The target region system based on Boost training and SVM classifier is designed.Before the proposed method,this paper also analyzed two methods based on connected domain and sliding window plus SVM classifier.The former method was unstable and greatly influenced by illumination,latter method has a high requirement for the parameter setting of the sliding window and is not intelligent enough.Through testing,this method can obtain high detection accuracy even in the face of images with complex backgrounds.