Research On Image Classification Method Based On Residual Cascade Network

Humans can easily identify different types of objects in the real world.For computers,before the computer is taught how to recognize objects in the real world through supervised learning methods and unsupervised learning methods,different types of images are same to those.There is no substantial difference,the only difference is that the order of "0" and "1" is different.In recent years,the rapid development of deep learning has made it relatively easy for computers to understand the real world.In deep learning methods,because convolutional neural networks have amazing results in image feature extraction and expression,convolutional neural networks are often used to achieve image classification.This article briefly introduces and analyzes the important research on image classification at home and abroad,including the traditional image recognition methodbag of words model,and the traditional image feature extraction methods involved,such as SIFT(scale non-transformation feature),LBP(Local Binary Mode),HOG(Histogram of Image Gradient),etc.;It also includes image classification network models that have a greater impact in deep learning methods,such as Alex Net,Google Net,VGG and Res Net.Traditional feature extraction methods focus on the extraction of detailed features,while convolutional neural networks focus on the extraction of overall image features.Both have their own advantages.Based on the improved Alex Net network and combined with traditional image feature extraction methods,this paper proposes a micro-expression classification method based on channel cascade to improve the accuracy of micro-expression classification.This method uses the MB-LBP operator and the SOBEL operator to perform preliminary feature extraction on the image,obtains the MB-LBP feature and the image gradient feature of the image,and then uses the improved Alex Net to construct a cascade network to extract features in parallel from the feature map and merge them.The final classification layer calculates the classification results based on the fusion features.In the development process of deep learning,many excellent image classification networks have emerged,among which Res Net has the greatest impact.This network first proposed the concept of "residual connection" to solve the problem of network performance degradation due to the increase in the number of network layers.This paper introduces the idea of "channel cascade" on the basis of "residual connection",and proposes image classification methods of residual connection and channel cascade network to improve the accuracy of network classification.The realization of this method relies on the convolution block containing "residual connection" and "channel grouping".Through the study of superimposing convolution blocks of different specifications,an RGNet model with higher image classification accuracy is obtained.This paper conducts experiments and analysis on the final model on the three data sets of FER2013,Ck+ and JAFFE-2.The experiment shows that the classification model has achieved a high recognition accuracy rate.