Feature fusion coefficient learnable image semantic segmentation method

Abstract

The invention relates to a feature fusion coefficient learnable image semantic segmentation method. The method mainly comprises the following steps: to begin with, training a deep convolution networkclassification model from image to category label in an image classification data set; converting a full connection layer type in the classification model into a convolutional layer type to obtain a full convolution deep neural network model for category prediction at the pixel level; expanding convolutional layer branch, and setting a coefficient for each branch, feature fusion layers being fusedaccording to coefficient proportion, and the coefficient being set in a learnable state; then, carrying out fine-tuning training in an image semantic segmentation data set, and meanwhile, carrying out coefficient learning to obtain a semantic segmentation model; carrying out fine-tuning training and fusion coefficient learning to obtain 1-20 groups of fusion coefficients; and finally, selecting the branch, the coefficient of which is largest, from each group, carrying out final combination, and carrying out fine-tuning training and coefficient learning again to obtain a final semantic segmentation model. The method enables the feature fusion effect to reach a best state.

Description

technical field [0001] The invention relates to the technical fields of deep learning and computer vision, in particular to an image semantic segmentation method with learnable feature fusion coefficients. Background technique [0002] In the past few years, deep convolutional neural networks have achieved great performance improvements in the field of computer vision, including image classification, object detection, pose estimation, and semantic segmentation tasks. At present, the main way to perform semantic segmentation tasks is to use deep neural networks for dense pixel prediction; at the same time, there are also some works that combine conditional random field methods to post-process the semantic segmentation results to make the segmentation results more refined. Previously, most of the semantic segmentation methods using deep neural networks first detected some object candidate boxes from the image, and then combined these boxes with categories as the segmentation r...

AI Technical Summary

Problems solved by technology

However, it is not that the more fusion features, the better. There is a lot of redundant information between semantic features. One-dimensional fusion of more features will increase the amount of calculation and calculation time, and may not help performance improvement; At the same time, there is no clear method for merging the features of which layers

Basically rely on intuition and experience to select layers, which is relatively blind

Images