A Multi-Instance Interactive Image Segmentation Method

Abstract

The invention discloses a multi-instance interactive image segmentation method, which can quickly segment a target image with multiple instances of arbitrary color distribution only by labeling a small number of pixel points, including steps: (1) acquiring an image containing manual labeling information ; (2) Carry out multi-instance mixed Gaussian modeling based on color information on the image; (3) combine the distance formula and multi-instance mixed Gaussian modeling information to classify the pixels in the image to obtain the segmented image; (4) The user can correct the segmented image and return to step (1); otherwise, save the image. Compared with the prior art, when solving the problem of multi-instance interactive image segmentation, the method of the present invention not only fully considers the color information of the image, but also considers the position information between the images, and realizes high-efficiency by comprehensively utilizing the color and position of the image. Performance of Image Segmentation.

Description

technical field [0001] The invention belongs to the technical field of image segmentation and fusion, and in particular relates to a multi-instance interactive image segmentation method. Background technique [0002] Image segmentation is to extract meaningful areas in the image through certain features (such as edges, textures, etc.) to determine whether there is an object of interest in the image; image segmentation is generally based on the similarity between pixels, pixel mutation or space distance etc. The so-called similarity between pixels is that the pixels in a certain area have similar characteristics, such as similar pixel values ​​and the same texture; A basis for segmentation, usually pixels with closer spatial distances are more likely to be classified into one category. There are many methods of image segmentation, which can be divided into semi-automatic segmentation and automatic segmentation according to whether manual participation in the segmentation pr...

AI Technical Summary

Problems solved by technology

The main disadvantage of the GrabCut method is that it can only segment images with a single foreground object, and requires the color distribution of the foreground and background pixels to meet the mixed Gaussian model, and requires a large difference in the color distribution of the two, which is not strong for the front and background contrast. The boundary area segmentation effect of the

Another popular method is the linear constrained spectral clustering method, which encodes the annotation information as a homogeneous linear equality constraint and adds it to the classic spectral clustering image segmentation framework to obtain interactive image segmentation results; this method does not Relying on the mixed Gaussian model of front and background color distribution, it can be applied to images of almost all scenes, but its disadvantages are: the calculation is time-consuming, it cannot be applied in real time, and the spatial smoothing information of pixels cannot be encoded into corresponding constraints. The utilization rate is low, and a large number of labeled pixels are required to obtain more accurate segmentation results

The interactive image segmentation method based on graph cut theory converts the target image to be segmented into a graph structure, and realizes image segmentation by converting the band solution problem into a problem of minimizing energy solution. Common methods include GraphCut, GrabCut and spectral clustering methods, etc., but there are two problems in the current interactive image segmentation algorithm based on graph cut theory: ① The number of extracted instances is limited; ② The segmentation algorithm is slow and cannot achieve real-time segmentation

[0004] The Chinese patent with the patent number CN 102360494A proposes a multi-foreground object interactive image segmentation method, which is different from the region-based interactive image segmentation. A discriminant analysis method is introduced in the field of the local window of the image, and the pixel is directly mapped to the category label through its feature vector; the disadvantage of this method is that it needs to estimate the error estimation of the category label of each pixel in the local window of the image, resulting in a large amount of calculation. And there may be similar or identical pixels within the local window, resulting in double counting

The Chinese patent with the patent number CN107730528A proposes an interactive image segmentation and fusion method based on the Grabcut algorithm. This method combines the Grabcut algorithm and the watershed algorithm to solve the problem of inaccurate segmentation of the similar front and background of Grabcut. However, the defect of this method is that Only limited to foreground and background segmentation

Images