Feature image extraction method based on deformable convolutional layer and feature image extraction device thereof

Abstract

The invention discloses a feature image extraction method based on a deformable convolutional layer. The method comprises the steps that a target image is acquired; and pixel values are extracted from the target image through the sampling points of the convolutional kernel in the convolutional layer so as to obtain a feature image, wherein the actual coordinate values of the sampling points are the actual coordinate values calculated according to the preset initial coordinate values and the deviation variable trained in advance. According to the method, one deviation variable can be additionally arranged for each sampling point in the convolutional kernel so that a convolutional neural network is enabled to realize the capacity of learning image space geometric deformation, the convolutional kernel is enabled to randomly sample near the present position, the adaptability of the convolutional layer for the deformed image in extracting the feature image can be increased and the identification accuracy for the deformed object can be increased. The invention also discloses a feature image extraction device based on a deformable convolutional neural network. The feature image extraction device also has the beneficial effects.

Description

technical field [0001] The present invention relates to the field of artificial intelligence, in particular to a method and device for extracting feature images based on deformable convolution layers. Background technique [0002] With the continuous advancement of science and technology, deep learning has achieved long-term development in recent years. At present, the most popular technology is convolutional neural network, and image detection based on convolutional neural network has become a basic research topic in various fields such as national defense and military, social security, public transportation, and commercial applications. The best image detection techniques to date have been implemented using convolutional neural networks. [0003] In the process of image detection, the same object may present different sizes, postures, perspective changes or even rigid deformation in the image, and the acquired image will be deformed for the same object. When the image is...

AI Technical Summary

Problems solved by technology

[0004] However, in the prior art, for a deformed object, a large number of feature images related to the object need to be stored, which will make the amount of data very large, which is inconvenient to store, and it will take a lot of time to compare the feature images

In the existing convolutional network architecture, it is impossible to adaptively adjust the receptive field of the convolution kernel according to the content of the image, which will limit the recognition accuracy of deformed objects

Images