A method for image foreground object segmentation

Abstract

The invention relates to a method for segmenting image foreground objects. Aiming at the characteristics of different areas of image foreground objects, a convolutional neural network for image foreground object segmentation based on boundary concerns is constructed. The network first uses the feature extraction backbone network to extract image features, and then uses the boundary The localization sub-network obtains boundary features and selective confidence maps, while the internal perception sub-network is used to obtain internal features and invariance confidence maps, and the transition supplementary sub-network is used to obtain transition supplementary features between the boundary and interior of the foreground object, and the output of the three-way sub-network The foreground object segmentation result is obtained through the feature mosaic selection method of boundary attention; next, the image foreground object segmentation convolutional neural network based on boundary attention is trained, and the image is input into the above-mentioned trained convolutional neural network to realize the image foreground object segmentation. The invention can effectively divide the foreground object into a whole, and at the same time, it can process the edge details well, and the image processing speed is fast.

Description

technical field [0001] The invention relates to the fields of computer vision and image content understanding, in particular to a method for segmenting image foreground objects. Background technique [0002] Image foreground object segmentation is an important basic problem in computer vision, and it is of great significance to tasks such as object recognition, target tracking, and image analysis. [0003] There are many existing image foreground object segmentation methods, and the traditional methods mainly consider using the comparison of global and local visual features to highlight foreground objects. In 2009, Achanta et al proposed using the color contrast of pixels to highlight foreground objects on CVPR; in 2011, Cheng et al proposed using the color histogram comparison of local areas to highlight foreground objects on CVPR; in 2012, proposed a method on CVPR to perform image superpixel segmentation in CIELAB space, and then use color and space to construct two met...

AI Technical Summary

Problems solved by technology

Another type of method mainly considers using other tasks or other clues to improve the performance of foreground object segmentation. In 2017, Chen et al. proposed on ICCV to use human gaze information and image semantic information to simulate the process of human manual labeling of foreground objects. In 2018, Chen et al. proposed a feedback mechanism that simulates the human cognitive process on ECCV to design a reverse attention supervision model for foreground object segmentation. This type of method usually focuses on the location of the foreground object and ignores the details of the foreground object.

Images