Image sorting method and system based on feature interaction and multi-task learning

Abstract

The invention discloses an image sorting method and a system based on feature interaction and multi-task learning. Among them, the image sorting method based on feature interaction and multi-task learning includes extracting the visual features of the original image; Using the extracted image visual features to carry out region-based image visual feature interaction; Multi-task learning neural network is used to aggregate the image visual features after interaction. The aggregated image visual features are input to the trained classifier for classification, and the images are sorted accordingto the classification results. Which has the effect of more accurate sorting results.

Description

technical field [0001] The invention belongs to the field of image processing, in particular to an image sorting method and system based on feature interaction and multi-task learning. Background technique [0002] The image generated by capturing the actual screen by input devices such as scanners and cameras is a bitmap composed of pixel dot matrix, which uses numbers to arbitrarily describe pixel points, intensity and color. Therefore, the image contains more features. At present, there are many methods for comparing and sorting various features of images, but they still face some problems: [0003] For example: when comparing two street view pictures which one is safer, more beautiful, more depressing, more lively, richer, and more boring, the existing method directly evaluates and scores the two pictures separately, and then compares them. The contrast between image regions is not considered, which can also be said to be a method of feature interaction, which cannot e...

AI Technical Summary

Problems solved by technology

[0003] For example: when comparing two street view pictures which one is safer, more beautiful, more depressing, more lively, richer, and more boring, the existing method directly evaluates and scores the two pictures separately, and then compares them. It does not consider the contrast between image regions, which can also be said to be a method of feature interaction. It cannot effectively notice the contrast between key elements in the image, which is not conducive to accurate judgment of two street view images.

Images