An image shape encoding method

Abstract

The invention discloses an image shape coding method, comprising the following steps: 1) extracting salient features of an original image to obtain a salient feature map; 2) binarizing that salient characteristic map to obtain a binary characteristic map; 3) shaping code of binary characteristic map to obtain shape code map. That image shape coding method of the invention can obtain the high recognition accuracy in the experiment after being applied to the CNN learn, which is generally higher than the highest recognition accuracy obtained when the method is not used. Therefore, the invention is helpful to improve the image recognition capability of the CNN.

Description

technical field [0001] The invention relates to image encoding in image processing, in particular to an image shape encoding method. Background technique [0002] Convolutional Neural Network (CNN) is an efficient intelligent image recognition method, which has achieved excellent results in practical applications in many fields. However, studies have shown that currently commonly used CNNs usually only learn low-level features (edges, textures) of images but fail to obtain high-level semantic features (shape, structure) that describe the global structure of images, which makes CNNs in image recognition. Sometimes there will be problems of misrecognition (recognizing two different images as the same object) and missing recognition (recognizing two similar images as different objects), which limits the highest recognition accuracy that CNN can achieve. Therefore, if the shape feature can be introduced into the learning of CNN, the recognition accuracy of CNN will be improved,...

AI Technical Summary

Problems solved by technology

[0003] The purpose of the present invention is to provide a method for encoding the shape of an image, and handing over the encoded image to CNN for learning can overcome the problem that it cannot obtain the advanced semantic features of the image in general

Images