Image Discriminative Region Extraction Method Based on Convolution Feature Spectrum

Abstract

The invention discloses a method for extracting discriminable regions of images based on convolutional feature spectrum, which includes acquiring pictures to be extracted and inputting them into a VGG16 convolutional neural network; The discriminant probability of the category; select the category of the maximum discriminant probability among all the discriminant probabilities of the picture to be extracted as the category of the picture to be extracted; the input feature spectrum calculation of the pooling layer of the kth convolutional stage in the VGG16 convolutional neural network can be Discrimination area: Bilinear interpolation is performed on the discriminable area obtained when k is 3 and the discriminable area obtained when k is 5, and the two discriminable areas are fused to form the final discriminable area of ​​the picture to be extracted.

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] The discriminable area refers to the target area in the image that is considered by the convolutional neural network as the basis for classification when the deep convolutional neural network is used to classify the image. When the convolutional neural network is applied to the task of image classification, features of different scales are extracted from the image, and these features are used for classification and discrimination. These features as classification and discrimination describe the target object. Extracting these feature information will be able to locate the discriminable target "pointed out" by the network, thereby extracting the discriminable area of ​​the image. Such a discriminable region spectrum intuitively represents the most promin...

AI Technical Summary

Problems solved by technology

This method has two disadvantages: 1. After replacing the fully connected layer, the classification ability of the convolutional neural network will be affected, and the accuracy rate is usually reduced by 2% to 4%. 2. Usually the network is input to the classification layer. The feature scales are very small and lack the edge information of the target object. The discriminable area spectrum extracted by using such features has insufficient grasp of the edge of the target object, making the discriminable area spectrum very blurred.

Images