Parallel multichannel convolutional neural network and construction method thereof, and image characteristic extraction method

Abstract

The invention discloses a parallel multichannel convolutional neural network and a construction method thereof, and an image characteristic extraction method and relates to the machine learning technology field. As depth sensors such as Kinect are widely applied, acquisition of a multi-mode image is made to be more convenient, so characteristic extraction research based on the multi-mode image information is of great importance. According to the method, a sub mode convolutional neural network model is established for each channel of the multi-mode image, and each mode depth characteristic vector is extracted; for acquiring multi-mode image characteristics having unified property, weight connection is established at a special full-connection layer by each mode sub network; in the full-connection portion of the multi-layer sub network, the multiple mode depth characteristic vectors are fused according to the weight ratio to form a fusion characteristic vector containing the information of each mode, and characteristic expression with lower dimensions and better expressive force can be acquired through multi-layer network training. The method is advantaged in that the characteristic expression can be applied to the identification and classification fields.

Description

Technical field [0001] The invention belongs to the field of machine learning, and more specifically, relates to a parallel multi-channel convolutional neural network and a construction method thereof, and a multi-modal image feature extraction method based on the parallel multi-channel convolutional neural network. Background technique [0002] In the field of computer vision and image processing, we can obtain the original representation information of the identified or classified object through measurement. This original characterization information can be obtained by direct measurement, so it is called the original feature, such as the gray value of each point in a digital image. Original features are easily perceivable by human intuition, but they are not often used in pattern recognition. The reasons for this mainly include three points: one is that the original features cannot reflect the essential characteristics of the object; the other is that the original features are...

AI Technical Summary

Problems solved by technology

[0005] Aiming at the above defects or improvement needs of the prior art, the present invention provides a RGB-D multi-modal image feature extraction method and system, thereby solving the problem that the current utilization of multi-modal information is mainly through the traditional extraction of each image form. The technical problem of not being able to obtain a truly effective fusion feature representation due to manual features

Images