Multi-scale image semantic segmentation method

Abstract

The invention discloses a multi-scale image semantic segmentation method. The method comprises the following steps: obtaining a to-be-segmented image and a corresponding label; constructing a full convolutional deep neural network, wherein the full convolutional deep neural network comprises a convolution module, a hole convolution module, a pyramid pooling module, a 1 * 1 * depth convolution layer and a deconvolution structure; setting hole convolution as channel-by-channel operation, and utilizing low-scale, medium-scale and high-scale characteristics in a targeted mode; training the full convolutional deep neural network, establishing a loss function, and determining parameters of the full convolutional deep neural network by training the sample image; and inputting the to-be-segmentedimage into the trained full convolutional deep neural network to obtain a semantic segmentation result. By means of the method, the image semantic segmentation problem with complex details, holes andlarge targets can be well solved while the calculated amount and the parameter number are reduced, and the consistency of category labels can be reserved while the target edges can be well segmented.

Description

technical field [0001] The invention relates to the field of computer vision, in particular to a multi-scale image semantic segmentation method. Background technique [0002] Semantic segmentation is an important issue in the field of computer vision. At present, various application scenarios (such as object detection, recognition, etc.) need to be realized through semantic segmentation. The task of semantic segmentation is to judge the category of each pixel of the image and label it. [0003] The semantic segmentation problem puts forward two requirements for the algorithm: (1) classification: the labels of the pixels in the target range belonging to the same category must be consistent; (2) localization: the pixels at the edge of the target can also be accurately classified. Among these two requirements, the former requires high-level semantic features, which can usually be obtained by setting the convolution step size, pooling downsampling, etc.; while the latter requir...

AI Technical Summary

Problems solved by technology

There are three deficiencies in DeepLab: (1) To a certain extent, it solves the contradiction between resolution and receptive field, but it is still not robust enough when segmenting targets of different scales; (2) DeepLab hollow convolution in each module The pixel at the position p of the feature image of the l-th layer is determined by the position k of the feature image p of the l-1 layer d *k d Neighborhood pixels calculated from

The same rate is used in the same module, which will cause a "holed" grid-like receptive field, such as image 3 (a)- image 3 (c), where image 3 (a) is the receptive field of the feature obtained after an ordinary 3×3 hole convolution with an expansion rate of 1; image 3 (b) is the receptive field of the feature obtained after two ordinary hole convolutions; image 3 (c) The receptive field feature utilization efficiency of the feature map obtained after three ordinary hole convolutions is low; although the deep feature image has a large receptive field range, the sampling of pixels in the receptive field is very sparse during calculation, and only a small part is used. pixel information, resulting in a serious loss of detail in the segmentation result; (3) as the number of dilated convolution modules increases, the pixel values ​​of the previous layer’s neighborhood with holes used when the new layer calculates a pixel, the values ​​of these pixels Has a large inconsistency, which is not conducive to the segmentation of complex shaped objects

Images