Training method and device for picture quality score model

Abstract

The invention relates to a training method and device for a picture quality score model. The training method comprises the following steps: obtaining a plurality of original samples, wherein the original samples comprise sample pictures and quality scores corresponding to the sample pictures; determining a plurality of pairing samples according to the quality scores corresponding to the sample pictures, wherein the pairing samples comprise paired sample pictures and quality comparison results corresponding to the paired sample pictures; training a double-path neural network based on the pairedsamples, and determining a single-path neural network in the double-path neural network as a feature extraction network, the feature extraction network being used for extracting feature vectors of the sample pictures; and training a preset machine learning model based on the feature vector of the sample picture to obtain the picture quality score model. The training method can train a picture quality score model with high accuracy based on a small number of samples.

Description

technical field [0001] The present disclosure relates to the technical field of image processing, in particular to a method and device for training a picture quality score model. Background technique [0002] In an e-commerce system, merchants often use a lot of pictures to display product information. Among them, a high-quality product picture needs to meet the following conditions: good-looking subjective intuition, high product recognition, prominent subject, and good quality of the picture itself. [0003] At present, when calculating the quality score of product pictures, the image quality score model is often used. In order to ensure the accuracy of the picture quality score, it is often necessary to use a large number of sample pictures to train the model, in order to train a high-accuracy model to calculate the quality score of the product picture. However, when training the picture quality score model, there may actually be only one or two thousand labeled samples...

AI Technical Summary

Problems solved by technology

However, when training the picture quality score model, there may actually be only one or two thousand labeled samples, and each sample corresponds to a quality score (0-100). At this time, if the ordinary deep neural network is used for direct training, there may be insufficient samples. problem, which leads to the failure of the trained quality score model to achieve the expected accuracy

Images