Research On The Key Technology And Application Of Multi-view Machine Vision Detection

Machine vision is an ideal non-contact inspection method,which is of great significance for improving production efficiency and product quality,and has broad application and development prospects.Image processing is one of the core technologies of machine vision.Whether accurate analysis results can be obtained from images determines the success or failure of machine detection.How to extract reasonable and effective features from images has always been a research hotspot in the field of image processing,and it is a key factor in determining the performance of the algorithm.The development of image feature representation method has laid a solid theoretical foundation for visual inspection.In addition,the effective use of multiple images information provides new ideas for improving the detection capability of machine vision.For example,enhancing the description ability of models or expanding the detection field of view.This paper focuses on the key problems of multi-view machine vision detection theory and application,and puts forward many new methods.This dissertation is divided into seven chapters,main research contents of each part are as follows:In chapter ?,the related study background and significance of this research are introduced.The development and application of machine vision are reviewed,and the research status of image feature representation theory,multi-view feature representation,image quality assessment,and image stitching are summarized.The main problems faced by research on multi-view machine vision detection are analyzed,and the main work of this thesis is also given.In chapter ?,the view-based feature description of 3D objects is studied,and a multi-view feature fusion method is proposed.Firstly,we employ the bag-of-words model to extract image features,and solve ambiguity during the process of quantization,and obtain a histogram of visual words frequencies as the feature of each view.Then,the overlap between views is estimated by calculating the similarity between them,and all views are merged to get a complete feature of the object.Finally,the importance of each visual word is considered when calculating model similarity.Compared with the existing methods,our method does not need to design complex camera arrays,and can acquire multiple view in a flexible way,and the integrated features are more reasonable in model matching.Experimental results of view-based 3D object retrieval(V3DOR)show that our method has higher performance.In chapter ?,the image quality assessment technology is studied.The characteristics of defocusing blur distortion caused by inaccurate focusing in machine vision system are analyzed.In addition,a no reference blur image quality assessment method is proposed.The algorithm extracts three statistical features of natural scene,frequency response and information entropy from patches of test image and one hundred natural pristine images,and learns a multivariate Gaussian model of test image and reference images,respectively.Then,the local distortion of the test image is estimated by calculating the distribution distance between the two models.Finally,a phase congruency map is used to weigh the local quality to calculate the overall quality of the image.The experimental results show that the algorithm has excellent performance and it can achieve evaluation results which are consistent with the subjective assessments,and the evaluation results are not affected by the specific content of the image.In addition,blur images of different types of detection objects are evaluated under machine vision scene,and the experimental results prove that the proposed method has good generality.In chapter ?,the image stitching technology is studied,and we proposed an image stitching algorithm which combines local optimized alignment,optimal seam-cutting and multi-resolution image fusion.Aiming at the problem that the images cannot be accurately registered because they do not satisfy the global homography,a local adjustment method is proposed to reduce projection error,and a global similarity constraint is employed to prevent image distortion.Generally,the existence of moving objects and parallax between images will make the overlapping regions not aligned accurately,resulting in ghosting and blurring.The optimal seam-cutting is searched in order to avoid directly blending contents of different images,but is might produce stitched seam.Then we expand the seam region and adopted the multi-resolution fusion method of Laplaian pyramid to improve fusion results.The proposed algorithm can effectively solve several kinds of problems during the stitching process,and can achieve better results than the existing methods.In chapter ?,the content-preserving image retargeting technology is studied,and an optimal bidirectional seam carving for compressibility-aware image retargeting algorithm is proposed.The algorithm is based on Seam Carving which is a kind of discrete retargeting operation.In order to overcome the shortcomings of existing image significance analysis methods,we put forward an improved energy function combining visual saliency map and unidirectional gradient map.In addition,we proposed a wall-seam model to evaluate the image compressibility and assign the right number of seams for each direction.We can retarget the image into a new aspect ratio by repeatedly carving out or inserting seams,and then scaling the stretched image to the target size.This algorithm can fit various types of images to display devices with different resolutions and aspect ratios.Compared with several state-of-the-art methods,our method introduces less visual distortion and preserves image content well.For the multi-view synthetic images appearing in some machine vision applications,the proposed algorithm can effectively retarget them into standard sizes,so that they can be displayed or processed normally.In chapter ?,aiming at the practical problems of visual object recognition,classification and dimensional measurement encountered in practical application,the corresponding machine vision inspection algorithms are designed,respectively.A two-stage template matching method based on local features is proposed to identify an irregular target and determine its scale level.The corresponding machine vision detection system has been applied to the real-time package sorting pipeline.In addition,a part image classification method that combines features such as color and shape is proposed to provide visual guidance information during the product assembly process.Based on the proposed algorithm,we build a machine vision part inspection system.Finally,a multi-view based machine vision dimensional measurement method is proposed.We use the image stitching technology to expand the effective field of view(FOV),which can achieve high-precision measurement of large-sized objects.In chapter ?,the major work of our research is summarized.We pointed out the innovations of this thesis,and also make a prospect for the subsequent research work.