Human visual perception-based image identification method

Abstract

The invention discloses a human vision-based image identification method. According to the method, an image identification structure capable of realizing cross-problem domain identification is established based on deep learning and human vision. By applying the model structure, images of a plurality of problem domains of a model can be identified, and a human vision system is further simulated. An original image is directly subjected to characteristic extraction by utilizing the property of human visual perception, namely, an HMAX method, so that heavy and complicated preprocessing steps are reduced and the calculation efficiency and feasibility of the method are improved. The number of parameters in the deep learning is reduced through an SDA (Stacked Denoising Autoencoder) method, so that the universality of an algorithm is enhanced and the training performance of general feed-forward BP is improved. An actual experimental result shows that the classification accuracy of the method is higher than that of other classification methods. Therefore, the method is an efficient and feasible image identification method, and has universal applicability in the field of image identification.

Description

technical field [0001] The invention relates to pattern recognition, artificial intelligence, computer vision, and stacked autoencoders. In particular, it relates to the object feature extraction model HMAX based on feature combination and the stacked autoencoder SDA under the deep learning model. Background technique [0002] Accurate image recognition has very important research significance. Image recognition technology plays an important role in many aspects such as medicine, aerospace, military, industry and agriculture. Most of the current image recognition methods use manual feature extraction, which is not only time-consuming and laborious, but also difficult to extract. Since the renaissance of deep learning, it has become a part of state-of-the-art systems in different disciplines, especially in computer vision. At present, the form of deep neural network has been proved to be almost the best structure in deep learning structure. [0003] Deep learning is a kind...

AI Technical Summary

Problems solved by technology

[0012] d) The same preprocessing method is not universal in different scenarios, resulting in low recognition accuracy of the same method in different scenarios

[0013] 2. The classic classification method based on human perception (eg, DBN) has too many training parameters

And the optimization process of the optimal result of an ultra-high-dimensional parameter is a rather complicated process.

This increases the complexity of the method used and reduces the computational performance of the algorithm

[0014] 3. Most of the current methods need to label the original image target area or background area

This process requires a large number of calculations and manual operations, which is not very practical

Images