Research On Scene Image Recognition Direction Based On Improved Convolutional Neural Network Algorithm

With the development of society,the artificial intelligence technology represented by deep learning has been widely used in scientific research and various aspects of daily life in recent years,such as face recognition,fingerprint recognition,illegal photography and smart car autonomous driving.Since much of human information comes from vision,image recognition is an important part of deep learning.How to perform image recognition is one of the key issues in the application of deep learning systems.Traditional image recognition separates image feature extraction and image classification,which requires artificial construction for image feature extraction.This not only increases the workload of people,but also greatly reduces the efficiency of image feature extraction.When faced with the feature extraction of some complex images,people tend to ignore the processing of most details,such as the image's color,texture,lightness and other shallow features,which also greatly limits the traditional image recognition application scenarios.In recent years,the application of convolutional neural networks has become popular.From the emergence of Alex Net convolutional neural networks in 2012 to the Mask-rcnn algorithm,convolutional neural networks have become practical.The biggest feature of convolutional neural networks is that feature extraction and classification are integrated into a single neural network.Convolutional neural network is a variant of multi-layer perceptron.It integrates the feature extraction function into the multi-layer perceptron by reorganizing the structure,reducing its own weight and eliminating the complex image feature extraction process before recognition.In the convolutional neural network,the close connection between layers makes it suitable for processing and understanding of images.At the same time,it can also automatically extract some rich related features from images.This article deeply studies the image recognition problem under the convolutional neural network.The main research contents include:(1)Improve the Alex Net network algorithm and apply it to large-scale image recognition.The traditional Alex Net convolutional neural network was improved by modifying the network framework,adding batch normalization algorithms,replacing the largest pooling kernel with the maximum mean pooling kernel and using the global mean pooling convolution kernel to replace the fully connected layer.Improve the original image recognition based on Alex Net convolutional neural network.It is undoubtedly important to improve the recognition accuracy in the field of image recognition,so the work in this chapter has certain practical significance.(2)In-depth study of multi-frame VGGNet convolutional neural network and applied to scene image recognition.The multi-frame VGGNet convolutional neural network model can retain the global information in the scene image on the one hand and the detailed information in the scene image on the other hand.Compared with the traditional single VGGNet convolutional neural network model,the recognition accuracy of scene images using this network model is higher and the extraction effect of feature images is also better.(3)Improved multi-frame VGGNet convolutional neural network algorithm and applied to indoor scene image recognition.First,the scene image is recognized based on the multi-frame VGGNet convolutional neural network and then the multi-frame VGG network framework is improved.By adding a batch normalization algorithm after each convolution kernel and using max-avg pooling kernels to replace the original max pooling kernels and using global average pooling convolutional layers instead of fully connected layers for feature extraction and finally using double The channel convolution technology performs coupling calculation on the double-frame network.When using the data set for training and testing,we used enhanced data to prevent the model from overfitting.The research in this paper realizes the effective combination of multi-frame convolutional neural network and batch normalization algorithm,which has a good practical effect on scene image recognition.