Saliency Region Segmentation Based On Multi-Feature Fusion And Conditional Random Filed

In computer vision,image segmentation is not only the basic part of image analysis,but also has a wide range of applications in object detection,image understanding,machine vision,automatic driving,medical image analysis and other fields,and has received increasing interest in recent years.The aim of image segmentation is to divide the image into meaningful pixel blocks,which simplifies the representation of the image and directly affect the subsequent high-level visual processing tasks.In the classical segmentation algorithms,Grab Cut has accurate segmentation result,but users need to not only draw the approximate area of the segmentation,but also specify the seed points in the background and foreground respectively.In practical applications,automatic image segmentation is required in certain situations.In this paper,the ultimate aim of algorithm is to achieve full-automatic image segmentation while ensuring segmentation accuracy.The algorithm mainly includes three main steps of preprocessing,feature extraction and feature fusion,and corresponding improvements are made as follows:1.Image preprocessing.In general,the algorithm takes the whole image as its input,and according to obseedges rvation,the boundary of the image is usually background area.Removing these areas not only does not affect subsequent significance calculations,but also reduces the complexity of subsequent algorithms.To solve this problem,this paper combines the superpixel SLIC block algorithm with the edge of the image,and regards the super-pixel block with insignificant as the background area and discards it,thus reducing the complexity of the algorithm.2.Feature extraction.Usually,the algorithm uses low-level features such as color contrast and spatial distance.Although the calculation is highly efficient,it does not simulate the vision system's process of analyzing the image well,and the result is relatively poor.In order to improve the performance of the algorithm,based on this,the paper introduces Local Binary Patterns(LBP),which is commonly used to calculate image textures;Region Contrast(RC)algorithm based on histogram acceleration to enhance the region Contact;and Dark Channel,which is mostly used for image defogging,but it has been experimentally found to reflect the integrity of brightly colored objects,thus ensuring the integrity of the salient regions.By introducing these intermediate features,the final segmentation accuracy of the algorithm is improved.3.Feature fusion and segmentation.The saliency maps extracted according to different features are fused in a linear manner.It is found that the final results vary with the set weights.For this problem,a conditional random field(CRF)framework is introduced to fuse different saliency maps..The traditional CRF calculates the weights of pixels in each pixel and eight neighborhoods to construct a grid-shaped graph,which is not only difficult to store,but also has a large amount of calculation.The CRF at the superpixel level proposed in this paper simplifies the algorithm and improves the final result.which performed.Finally,the proposed image segmentation algorithm is tested on the MSRA-1000 image database.The experimental results are compared with not only classic algorithms such as FT and RC,but also DSR,HS,MC,MNP,SEG,GR and GMR methods appear in recent years.From the results,the Precision,Recall and F-Measure of the algorithm have better performance.