Direction superpixel-based rapid image segmentation method

Abstract

The invention discloses a direction superpixel-based rapid image segmentation method. Compared with the segmentation performance of traditional segmentation methods based on clustering, watersheds, active contour models or graph models, the segmentation performance of the method is averagely improved by 100%. Compared with some previous segmentation methods which predict edges through a convolutional neural network and adopt time-consuming post-processing, the method can operate in real time with the speed being more than 18 times that of the segmentation methods. According to the segmentationmethod of the invention, a two-dimensional vector is predicted at each pixel position through a convolutional neural network, and the direction of the vector points to a current point from an edge closest to a current pixel; a direction-based superpixel graph is obtained according to the predicted direction of each position; on the basis of the superpixel graph, a regional relation graph is constructed; and a segmentation result is obtained by using a customized rapid fusion method. The method achieves a very good effect on the balance of the speed and precision of image segmentation, is simple to implement, and has a very wide practical application range.

Description

technical field [0001] The invention belongs to the technical field of computer vision, and more particularly, relates to a fast image segmentation method based on directional superpixels. Background technique [0002] Convolutional neural networks have greatly improved the performance of various computer vision tasks, such as image classification, object detection, semantic segmentation, object tracking, etc. The goal of semantic segmentation is to assign a semantic label to each pixel in the image. Although the accuracy of semantic segmentation is very high at present, it is difficult for a trained model to obtain accurate segmentation results for unseen scenes or categories. General image segmentation is different from semantic segmentation. Its purpose is to divide an image into several non-overlapping regions, and maintain semantic or visual perceptual consistency within each region. [0003] Due to the inability to determine the number of blocks for each image segment...

AI Technical Summary

Problems solved by technology

[0003] Since the number of blocks for each image segmentation cannot be determined, the current high-precision or fast semantic segmentation model cannot be migrated to general image segmentation, and the existing general image segmentation methods can be roughly divided into high precision but slow speed and low precision but fast speed two types

High-precision image segmentation methods generally use convolutional neural networks to predict edges, and then use very time-consuming watershed methods to obtain segmentation results. The total time is close to 1 second. Although the accuracy is high, the speed limits its practical application.

The fast method uses the convolutional neural network to predict an embedded space, and then uses the clustering method to obtain the segmentation result. This type of method will cause serious leakage problems at the weak edges in the image, resulting in low accuracy, and the speed is far from real-time operation. The difference is far, and the practical application is very limited

Images