Research On Transparent Object Detection And Segmentation Algorithm Based On Light Field Information

Transparent object detection and segmentation is a hot and difficult problem in the field of machine vision.The analysis of an imaging scene containing a transparent body becomes very complicated due to the interaction between a transparent object and light.Unlike Lambertian body imaging,transparent body has no image itself,and its appearance is essentially a distorted view of the background.Even slight changes in the perspective may cause significant changes in the image.Conventional machine vision algorithms based on Lambertian hypothesis cannot effectively detect and segment transparent objects in scenes.Light field imaging technology is widely used in machine vision research because it can obtain more scene information and is beneficial to subsequent visual algorithm processing.Therefore,based on light field imaging theory,this paper uses light field information and visual features of transparent targets to carry out the detection and segmentation method research of transparent targets.By referring to the relevant literature at home and abroad,based on understanding the research status of relevant issues,and according to the visual characteristics of transparent target imaging and light field imaging theory,the feasibility of using light field information to realize the detection and segmentation of transparent target is analyzed.Aiming at the problem that traditional machine vision algorithm is difficult to detect transparent objects in scene,this paper uses four-dimensional light field plane consistency to detect transparent objects in scene.In order to improve detection efficiency,this paper proposes a fast image matching algorithm based on feature saliency and adaptive density clustering.The reliable feature guided image was used for triangular region segmentation,and the regional significance of image features and triangular region correspondence between multi-view images were used to improve the matching efficiency of multi-view images and achieve rapid and efficient detection of transparent targets.Traditional monocular camera can not effectively perceive the visual distortion characteristics of transparent target,and optical field imaging can be used to extract the visual characteristics of transparent target.Based on the transparent object detection,a transparent object segmentation algorithm based on the fusion of light field information and edge perception was proposed to solve the problems of incomplete transparent object segmentation and imprecise boundary segmentation based on visual method.In this algorithm,the edge perception method of transparent target and the consistency estimation method of light field plane are optimized by the method of light field sub-view region segmentation.The edge perception ability of transparent target is improved,and the influence of error introduced by mismatching on the segmentation result of transparent target is effectively suppressed.The energy function of transparent target segmentation was constructed by using light field plane consistency estimation and boundary constraint,and the segmentation precision of transparent target in scene was improved by combining with graph cut method.Although light field imaging can perceive the imaging features of transparent targets,the visual features of transparent targets cannot be effectively extracted when the parallax pole between the sub-view images of narrow baseline light field image is small.To solve the problem that the transparent target is not obvious in the narrow baseline and small parallax light field image,a transparent target segmentation algorithm based on the four-dimensional plane consistency of the light field superpixel is proposed.In this algorithm,the superpixel of light field is used as the processing unit to accumulate the effect of transparent target on the light corresponding to the superpixel of light field and amplify the visual features of transparent target,which is beneficial to the segmentation of transparent target in narrow baseline and small parallax light field image.The energy function of transparent target segmentation is constructed by integrating the four-dimensional plane consistency estimation of the light field superpixel and the autocorrelation of the light field superpixel,and the image cutting method is combined to achieve the segmentation of transparent target in narrow baseline light field image.The main research content of this paper is the detection and segmentation of transparent targets.Experiments are carried out using public data sets and self-collected data to verify the feasibility of the transparent target detection and segmentation algorithm based on light field information.The experimental results show that the proposed transparent target detection algorithm is faster than the existing similar algorithms.The detection speed is significantly improved without reducing the detection accuracy,and the detection speed is increased nearly5 times.In this paper,the transparent object segmentation algorithm based on the fusion of light field information and edge perception improves the integrity and precision of transparent object segmentation.Narrow small baseline parallax image of a light field difficult for transparent target segmentation problem,the paper put forward based on 4 d light field super pixels in the plane of consistency constraints of narrow light field baseline transparent target image segmentation algorithm is effective in achieving this kind of light field image transparent target segmentation,comparing algorithms,the thesis puts forward the algorithm of segmentation accuracy received a significant boost,g Acc increased by 0.0414,m Acc increased by 0.4387,m Io U increased by 0.4986,w Io U increased by 0.1281,m BFS increased by 0.3376.