Research And Application Of Deep Convolutional Neural Network Model Based On Multi-region Features

Deep learning has made great progress in the field of artificial intelligence and deep learning has a wide range of applications in face recognition,object detection,and speech recognition.In the field of deep learning,image classification by a convolutional neural network can directly perform a convolution operation on the image,thereby directly extracting the feature information of the image.Convolutional neural networks are considered to be one of the best techniques for learning image content,and have shown the best results in image recognition,segmentation,detection,and retrieval-related tasks.However,most convolutional neural networks can only perform one-way training and learning for one of the image channels,and can only use the entire image as input to extract features while ignoring the features of multiple regions of the image as input,which makes some important images Information is inevitably lost.Existing deep convolutional neural network models also have problems such as complex network structures and large calculations,so they cannot be widely used in practice.This paper mainly analyzes and studies from the perspective of improving the deep convolutional neural network model and the combination of the deep convolutional neural network model and the lightweight network structure.And then proposes a deep convolutional neural network model(MR-CNN)based on multi-region features and improved MRSS-CNN model based on SE module.The main content of this paper is as follows:1?The development process,basic structure,innovation,improvement of four deep convolutional neural networks and four classic convolutional neural network models are introduced.2 ? A deep convolutional neural network model based on multi-region features is proposed(MR-CNN).The model firstly divides the image into multiple regions,and then use standard convolution operation to get the feature information of the image,then use multi-region input to learn contextual interaction features.By cascading the feature information of the global region and multiple sub-regions to obtain more useful featureinformation,and then input the cascaded feature information into the convolutional layer,so that the model extracts the contextual feature information of the image in a supplementary way and finally the images are classified by the Softmax function.Experiment on the MNIST dataset and the Cifar-10 dataset,the experimental results show that the model has a simple structure,less parameters,and multi-region feature fusion context information modeling has better robustness and higher classification accuracy than single-region feature modeling.3? In order to further improve the learning ability of MR-CNN,this chapter studies the structural characteristics of SENet lightweight network and integrates the SE block structure into the MR-CNN model.The improved MRSS-CNN model not only takes into account the multiple regions of the image,but also takes into account the importance of the feature channels in different regions of the image and the use of jump connections to fuse the underlying features with higher-level features.The experiments on Cifar-10 dataset and STL-10 dataset verify that the MRSS-CNN model can further improve the classification effect and robustness of the model.4 ? This paper applies MR-CNN model and MRSS-CNN model to specific image classification tasks.Experimental results show that both the MR-CNN model and the MRSS-CNN model have achieved good results in the task of garbage classification,but the MRSS-CNN model has higher accuracy and robustness than the MR-CNN model,and has more application value.