An image quality evaluation method based on importance between deep network features

Abstract

The invention discloses an image quality evaluation method based on importance between deep network features. According to the method, a module for judging the importance relationship between the feature maps is added to end-to-end neural network model training, the image quality can be predicted more accurately, and a high generalization capability is shown on each image quality assessment (IQA)data set. The method specifically comprises the following steps of 1) preparing an image quality assessment data set for training a test network model, and randomly dividing the image quality assessment data set into a training set and a test set according to image contents; 2) adding a SeNet module into the vgg-16 network to build a neural network model VGG*-SE in multiple different combination mechanisms for image quality assessment, using the training data set to train the network models respectively, selecting the model as a final model when the trained model reaches expected precision onthe test data set, and storing parameters after training of the network model; and 3) calculating the prediction precision of the test set by using the selected final model.

Description

technical field [0001] The invention belongs to the field of computer image processing, in particular to an image quality evaluation method based on the importance among deep network features. Background technique [0002] As the basic content in the field of image processing, image quality assessment has a wide range of practicability in many fields, such as image compression, video codec, video surveillance and so on. Since it is difficult to obtain reference images of images to be evaluated in many practical applications, it is becoming increasingly important to develop effective no-reference image quality assessment (NR-IQA) methods; NR-IQA methods mainly include traditional methods and deep learning methods . Traditional methods mainly use low-level features manually extracted and related to human perception, and evaluate image quality through shallow regression methods. The main disadvantage of this method is that the manually extracted low-level features are not eno...

AI Technical Summary

Problems solved by technology

This network introduces the SeNet network used in the image classification field into the IQA problem, making it easy to distinguish the difference between the different channel features in the convolutional layer caused by sending the image into the IQA network

Images