An image classification and recognition method for depth feature estimation based on subspace decomposition

Abstract

The invention relates to an image classification and recognition method for depth feature estimation based on subspace decomposition. The method comprises the steps of S1, obtaining an unshielded image and an additional image pair of a training set; S2, extracting a depth feature vector from each input image; S3, calculating the difference between the depth feature vectors of the occluded image and the depth feature vectors of the unoccluded image in the additional image pair to obtain an error vector; S4, forming a first subspace by using the calculated depth feature vectors of all the unshielded images, forming a second subspace by using the error vectors, and forming a cascade dictionary based on the first subspace and the second subspace; S5, calculating a coefficient matrix based on the cascade dictionary; S6, training a classifier based on all column vectors of the first subspace; And S7, performing classification and identification on the to-be-identified image based on the cascade dictionary, the coefficient matrix and the classifier. Compared with the prior art, the method has the advantages of wide application range and the like.

Description

technical field [0001] The invention relates to an image classification and recognition method, in particular to an image classification and recognition method based on subspace decomposition for depth feature estimation. Background technique [0002] Occlusion happens from time to time in real life. Classifying occluded images is not difficult for the human visual system. However, in the field of computer vision, it remains an extremely challenging task even for deep convolutional neural networks that have recently achieved great success in many computer vision tasks. State-of-the-art convolutional neural networks have tens of millions of parameters, so even classification of unoccluded images usually requires a large data set to support good results. However, in practical applications, it is undoubtedly very difficult to collect a large amount of occlusion image data. A popular option is to train the network directly on datasets of unoccluded images or datasets containi...

AI Technical Summary

Problems solved by technology

First, these works only improved the results for face recognition, but did not show good results for occluded images in general.

Second, these jobs are not suitable for future cloud-based applications because they attempt to restore

Third, these works usually require a large number of partially occluded or missing images and time-consuming training of generative models

Finally, these efforts require retraining or fine-tuning the generative model for new occlusion patterns, which is often complex and time-consuming

Images