A Method of Object Classification and Pose Detection Based on Deep Convolutional Neural Network

Abstract

The invention belongs to the field of image processing and target detection, and discloses a target classification and attitude detection method based on a deep convolutional neural network. The core content of the method includes: 1. a method for making a sample set; 2. a method for generating candidate windows; 3. a method for feature extraction; 4. a method for network training; After the image to be detected gets the deep convolution feature through the convolutional neural network, the candidate window with the attitude angle is mapped to the feature layer to obtain the directional regional feature vector, and the final detection result is obtained by classifying and predicting the feature vector . This method can extract more pure target features from samples, improve the accuracy of classification, and realize the function of target attitude angle detection.

Description

technical field [0001] The invention belongs to the field of image processing and target detection, and discloses a target classification and attitude detection method based on a deep convolutional neural network. Background technique [0002] Object detection needs to accurately mark the location of the object on a given image and identify the type of object. The variable size of the target, the variable position of the target in the image, the variable attitude angle of the target, and the change of background illumination will cause difficulties in the detection process, resulting in a decrease in the detection accuracy. [0003] Classical target detection methods usually use template matching plus sliding window to solve the problem of target recognition and positioning in the image, which takes a long time and is less effective when the target changes drastically. The emergence of convolutional neural networks has led to the development of this topic, which has greatly...

AI Technical Summary

Problems solved by technology

Therefore, if the target framed by the upright rectangular frame will contain a lot of background information, the signal-to-noise ratio will drop significantly, which is not conducive to the subsequent target classification and position correction.

And because the method does not take into account the target's attitude angle changes in the image, it cannot characterize the target's pose

Based on the above points of view, the current target detection methods have limited detection capabilities, and the representation of the target state is not sufficient.

Images