Research On Image Features Matching And Application Of CNC-Vision System

Since1990s, the rapid development of computer technology has been improving CNC technology. The developing trend of CNC with equipments points to the director of multi-technology integration and adaptive ability. If computer vision can be applied to CNC machine, it will make a big step in this director.Applications of computer vision onto CNC machine include assessing tool wear, evaluating roughness of workpiece, planning tool path, security monitoring and so on. All these applications are related to image registration, which is a basic study aspect in image processing and analysis and performs image analysis, change detected,3D reconstruction, object recognition and image retrieval and processing. So image registration plays an important role in application of computer vision onto CNC machine.Registration based on image features has been a hot topic as it produces high precision, high speed and stability. The registration is done with features finding and matching, and keypoint of it is how to describe the features to achieve accurately fast matching. In recent years, lots of descriptor algorithms have been emerged, but neither of them can get high accuracy and real-time performance at same time. This paper focuses on that how to build better descriptors to improve registration with high performance and high speed from different aspects, and that how to apply such registration to reverse manufacturing and security monitoring.(1) In order to improve the robust of binary descriptor based on general sample pattern, a new describing feature method based on rank-ordered sample pattern is proposed. The sampling pattern constituted by patch, sub-patch and sampling-points corresponds to a descriptor constituted by bit string, bit segments and bits, and each sub-patch and each sampling point is ordered arrangement. Gauss-Laplace (LOG) is used to process each sub patch, and the value of each sampling point is determined according to the positive and negative of the LOG response. Experiment results show that the proposed descriptors can be match not only accurately but also fast in various images transforms of zoom, rotation, blurring, illumination varying as well as smaller viewpoint changes.(2) From pixel clustering perspective, a feature describing method based on general sample patterns in which sample points are rank ordered is proposed. Firstly, several sub-patches could be obtained in a way of decomposing one original image patch by sorting all pixels on it according to their intensity order, then random tests are token on each sub-patch, and concatenating all test results to form a distinct binary string of sub-patch; Secondly, multiple original image patches can be got by dividing interest region into several patches of different size around the keypoint, or can be retrieved from scale space of image patches, and the discriminative power of descriptor could be further raised by taking tests on these multiple patches. It may draw a conclusion that excavating feature information in local image region could be benefit to improve the robust of binary descriptor encoding not only intensity-comparison information but also information about relative relationship of intensities.(3) From view of machine learning, a method to learn low-dimensionality and high-discriminative descriptor based on AdaBoost framework is proposed. In the framework, the representations of image patches are modeled by non-linear weak learners which are trained through Adaboost algorithm. A similarity function is proposed to use as kernel function of optimized object function to learn a similarity embedding. The proposed learning framework is more generalizing than others as it may be not restricted to any predefined feature-sampling model, so it could encompasses intensity and director-gradient information. The results show that the ability to effectively optimize over the descriptor filter configuration leads to a significant performance boost of the feature descriptor with strong generalization is high robust to image transformations.(4) According to the characteristics of CNC image, and the requirement of CNC-Vision system that iamge feaure matching is high robust to changes of perspective and scale, two kinds of sample pattern are proposed in tandem, which combined with ordered rank, pixel clustering or integration learning to construct a image feature descriptor for CNC-Vision system. Through further analyzing sampling-pattern characteristics of BRISK and FREAK, it finds that both the sampling-point density and the degree of overlapping have an influence on the specificity of descriptor. It could appropriately tune the two factors to design an optimized sampling pattern, and map it onto the local area of keypoint with right orientation. A coarse descriptor, built by testing sampling points selected randomly on the sampling pattern, can be used to learn a fine descriptor from training data. Results based on experiments of performance evaluation under two kinds testing environments have shown that the proposed binary descriptor outperforms others. The good effectiveness of applying the proposed descriptor into application of3D construction has proved that the proposed descriptor is robust to variety of image transformations, as well as performs well in real-time applications. And then a more robust binary descriptor is proposed through further excavating feature information of image patch. Conventional binary descriptors such as BRIEF are not to rotation and viewpoint invariance, which is improved from two aspects in this paper. Firstly, an optimized sampling pattern is presented by tuning the density of sampling points and the overlapping size of receptive fields. Secondly, all pixels in the patch are classified according to their intensity order, so that the patch is decomposed into several sub-patches. Then it needs to repeatedly take random tests on each sub-patch mapped with the optimized sampling pattern, and concatenate each test result to form a distinct binary string of sub-patch. The proposed descriptor encodes not only intensity-comparison information but also information about relative relationship of intensities. Results based on experiments of performance evaluation have shown that the proposed binary descriptor outperforms art-of-the-state binary descriptors and is able to complete feature matching in CNC-Vision system.(5) Research on application of image registration based on image feature matching in CNC-Vision system. Firstly, it introduces multi-view stereo technology into reverse manufacturing and describes the improvement of densification and refining for matching point base on PMVS and proposing a method of expandable duplicating densification of space point cloud for the matching point to produce a more pronounced texture3D model; Also it proves the feasibility of reverse manufacturing remodeling using the technical of the visual three-dimesnsional reconstruction. Secondly, a method based on computer vision to monitor CNC machining in security, and ensure live safety of operators is proposed, which is achieved by zoning frames into different areas, constructing background model for each pixel in all the areas, and sending passages to control CNC machine according to inconsistence in foreground-background. It could be used to protect personal safty as this vision security monitoring well adapt lighting changes and machine vibration.