Image segmentation model training method and device, image segmentation method and device, equipment and medium

Abstract

The invention provides an image segmentation model training method and device, an image segmentation method and device, equipment and a medium, and the training method comprises the steps: obtaining asample data set which comprises a plurality of training images, wherein each training image is marked with a corresponding segmentation label; and training a pre-established image segmentation modelaccording to the sample data set, the image segmentation model comprising an encoder and a decoder, and the decoder performing up-sampling by using a sub-pixel convolution network. According to the invention, sub-pixel convolution is used to replace interpolation operation or transposed convolution operation commonly used in a decoder of an image segmentation model to carry out up-sampling, the problem that too many artificial errors are introduced into traditional interpolation operation or transposed convolution operation can be solved, and the image segmentation precision is improved.

Description

technical field [0001] The present invention relates to the field of image processing, in particular to an image segmentation model training method, image segmentation method, device, equipment and medium. Background technique [0002] The image semantic segmentation task refers to assigning a semantic label to each pixel in the image. The early semantic segmentation algorithm used FCN (full convolutional network), which is an end-to-end convolutional neural network architecture. This algorithm can generate segmentation maps for images of any size, but the existing problems are: in The convolution operation on the original image resolution is expensive. In order to solve this problem, FCN adopts down-sampling and up-sampling processing, but due to operations such as pooling, a large amount of information will be lost, causing FCN to generate a rough segmentation map. In order to obtain a more efficient segmentation effect, a U-shaped framework including an encoder and a de...

AI Technical Summary

Problems solved by technology

The early semantic segmentation algorithm used FCN (full convolutional network), which is an end-to-end convolutional neural network architecture. This algorithm can generate segmentation maps for images of any size, but the existing problems are: in The convolution operation on the original image resolution is expensive

The traditional interpolation method is the simplest and commonly used method. It directly bilinearly upsamples the feature map to a specified multiple, but its disadvantage is that it is too simple and has no parameters to learn, and it is easy to introduce artificial errors. If the upsampling If there are many multiples, the segmentation effect will be poor; the advantage of transposed convolution is that it can be trained, but the disadvantage is that there is zero padding operation, and it is easy to introduce artificial errors, the effect is not very ideal, and it is more difficult to train

Therefore, regardless of whether the decoder of the image segmentation model adopts interpolation operation or transposed convolution operation for upsampling, satisfactory image segmentation effect cannot be achieved.

Images