Intelligent segmentation method integrating target detection and image segmentation

Abstract

The invention discloses an intelligent segmentation method integrating target detection and image segmentation. According to the method, a target is extracted by utilizing a high-dimension feature extracted by a neutral network; and a deep learning-based target detection network Faster-RCNN has high efficiency and can quickly and accurately identify a target region, so that a target region of interest can be extracted firstly by utilizing the characteristic and then the region is subjected to target segmentation pointedly. The method can automatically detect and extract the region of interest, reduce occupation of calculation resources of a GPU, and quickly and accurately perform segmentation to obtain the target; and under the condition of a huge image quantity scale, the method has efficient implementation performance, so that the manual interaction process can be reduced.

Description

technical field [0001] The invention relates to computer graphics and image processing technology, especially for the intelligent segmentation of target object sequence images in three-dimensional reconstruction, specifically an intelligent segmentation method integrating target detection and image segmentation. Background technique [0002] In 3D reconstruction based on image sequences, the accuracy and speed of object segmentation play a decisive role in the quality and efficiency of reconstruction. Image segmentation is widely used in military, remote sensing, meteorology, medicine and other fields, and it is a difficult point in image processing. Due to the large number of images, the extremely complex and changeable surrounding environment, and the variety of object materials, it is impossible to use a single method for batch automatic segmentation. Currently, Photoshop software is used for target segmentation based on image sequences, and the process of manual particip...

AI Technical Summary

Problems solved by technology

For the target object in 3D reconstruction, the gray value information is considered separately, and a large amount of image color texture information is discarded, resulting in extreme sensitivity to noise, and it is impossible to find an accurate threshold dividing line, so the target cannot be well separated from the background. separate

[0004] 2) The edge detection segmentation method is to determine the boundaries between different regions according to the drastic changes in the pixel values, because the images in the same region have the same texture and color distribution, so the place where the first derivative or second derivative of the pixel is detected suddenly That is, it is marked as an edge, but this method also has limitations. There are a large number of large and small edges in the entire image, and not all edges can be connected to form a closed area, so the target area cannot be accurately determined

However, the direct watershed algorithm based on the gradient image can easily lead to over-segmentation of the image. The main reason for this phenomenon is that there are too many small areas in the input image and many small catchment basins are generated, which leads to the segmentation of the image. Represent meaningful regions in the image

Images