Image discriminable region joint extraction method based on convolution characteristic spectrums of multiple groups of k classifications

Abstract

The invention discloses an image discriminable region joint extraction method based on convolution characteristic spectrums of multiple groups of k classifications, and belongs to the field of image processing. The method comprises the steps of randomly selecting k-1 classes and a current class to form a group for any class based on picture types of a training set, and respectively selecting a group a for each class; respectively inputting each group of pictures into a preset convolutional neural network for convolutional neural network training to obtain a plurality of trained convolutional neural networks; inputting the to-be-extracted picture and the image-level label thereof into a number of a trained convolutional neural networks matched with the picture type; and for each matched trained convolutional neural network, obtaining the convolutional neural network through weighted gradient class activation mapping to generate a group of discriminable regions, and fusing the number ofa groups of discriminable regions to obtain a final discriminable region of the to-be-extracted picture. According to the invention, the more complete discriminable area of the image can be extracted,and the integrity of the extraction result of the existing extraction method is improved.

Description

technical field [0001] The invention belongs to the field of image processing, and in particular relates to an image discriminable region extraction method based on convolution feature spectrum. Background technique [0002] Discriminative region generation for convolutional neural network classification models is a fundamental step in weakly supervised analysis for computer vision and can be applied to tasks such as object localization, segmentation, and recognition. The discriminable area refers to the area activated by the convolutional neural network in the image when the deep convolutional neural network is used to classify the image, which is considered as the target area for classification. It is an effective way to transfer from image-level weak labels to object regions. At the same time, extracting discriminable regions of images is conducive to in-depth understanding and visualization of deep convolutional neural networks, which plays an important role in the rese...

AI Technical Summary

Problems solved by technology

[0003] The current algorithm for image discriminable area extraction based on deep convolutional neural network is to extract discriminable areas based on all categories of images. Since this method classifies all categories, the most critical discriminable area is obtained. That is, the discriminable area is relatively small, and it is impossible to extract a complete discriminable area

Images