Semantic segmentation method based on multi-scale deformable convolution

Abstract

The invention relates to a semantic segmentation method based on multi-scale deformable convolution. The semantic segmentation method comprises the following steps: step 1, preparing semantic segmentation input image data for training and corresponding labels thereof; step 2, training the deep learning network; firstly, a basic model of a network is pre-trained; parameters of the network are further optimized on the basis of pre-training; a relevant semantic segmentation module is added to carry out further training to finally obtain network parameters suitable for the data set, and for a newly input image, the network carries out forward calculation to finally obtain the output of the image, so that the image can classify each pixel in the image to form a semantically segmented output image; and step 3, carrying out weight updating on the loss function by adopting a gradient descent method, so that the loss is gradually reduced, and carrying out iterative training until the network converges or reaches the maximum number of iterations to obtain final network parameters, and storing the trained network model and each parameter weight to form a semantic segmentation model.

Description

technical field [0001] The invention belongs to the field of semantic segmentation and relates to a method for semantically segmenting images by using multi-scale deformable convolution. Background technique [0002] Semantic segmentation is the task of classifying pixels into the categories identified by the dataset. It is a fundamental and challenging field in image processing. This technology is widely used in different fields such as self-driving cars, disease detection in medical images, and drone flight experiments. [0003] In recent years, as the most basic and instructive network at present, based on the fully convolutional neural network [1] (FCN) has achieved a major breakthrough in semantic segmentation methods. Different from the classification network, FCN predicts the pixel category by replacing the fully connected layer of the classification network with a convolutional layer. However, this method predicts each pixel based on a small range of images, whic...

AI Technical Summary

Problems solved by technology

So if you just use multiple parallel expansion convolutions in the final output of the basic network like Deeplabv2, you can't extract spatial information very well.

Images