End-to-end difference network learning method for image semantic segmentation

Abstract

The invention discloses an end-to-end difference network learning method for image semantic segmentation. The method comprises the steps that a main network structure and a complete network structureare constructed respectively by using a Caffe deep learning framework, wherein the main network structure is used to generate a rough segmentation model and a small target area of each image in a training set and the complete network is used for final image semantic segmentation; the rough model of the main network is trained by using a part of the data of the training set, and a segmentation result acquired by the rough model is compared with a real segmentation image to acquire a mistaken segmentation area of the rough model; the acquired rough model is used as an initialization parameter totrain a complete network model to acquire a final segmentation result, and an image semantic segmentation model is established; the segmentation model is tested; and all test images are segmented according to the image semantic segmentation model acquired in the step 3. The method can be sensitive to small target areas, and can solve the problems of edge blur and misjudgment of similar parts to some extent.

Description

technical field [0001] The invention belongs to the technical field of image processing, in particular to an end-to-end difference network learning method for image semantic segmentation. Background technique [0002] Image semantic segmentation is one of the three major tasks of computer vision. Its goal is to classify each pixel in the image and obtain a semantic segmentation map of the image. From the perspective of traditional image segmentation, image semantic segmentation is to segment the image into multiple regions at the semantic level, and then assign appropriate category labels to each region. At present, semantic segmentation has a wide range of applications in autonomous driving, real-time road monitoring, automatic virtual fitting, and medical disease systems. Before the rise of deep learning, the main method of semantic segmentation was to use the conditional random field model to build a probability graph model. In recent years, due to the strong learning ab...

AI Technical Summary

Problems solved by technology

[0004] ① Difficult to segment small target areas or strips of target areas

[0005] ②Difficult to segment blurred border areas

In complex natural scenes, both foreground and background regions contain more than a single object, which poses a great challenge to the task of semantic segmentation

At the same time, the segmentation performance at the boundary of two objects is affected by similar colors, blurred edges, mutual occlusion and illumination changes. The existing segmentation methods can only get the main part of the target and cannot get the precise edge of the target.

[0006] ③ Difficult to segment similar parts of different targets

At the same time, it is difficult for a single model to handle these boundary problems well without reducing the overall accuracy

Images