Design And Implementation Of A Customized Image Recognition System Based On Deep Learning

Recently,deep learning technology has achieved great success in the field of artificial intelligence,especially in image,speech recognition and language processing to show great advantages of use.And image recognition based on deep learning has been widely used.But these applications are currently existing in some larger enterprises or organizations,and it is more difficult for smaller companies,some individuals and teams want to customize an image recognition model according to their needs.Therefore,in order to reduce the complexity of people to customize image recognition models,this thesis designs and implements a custom image recognition system based on deep learning.This system only requires the user to create the dataset according to their own needs,then upload it to the system,and select model performance in the system customization page.The system will train the user's dataset,evaluate the model and deploy the model,thus achieve the customization of image recognition models.The main work of this thesis is as follows:(1)Research on customized image recognition model.When customizing the image recognition model,it can be customized in three dimensions as follows:Firstly,the three core models of image recognition,namely image classification,semantic segmentation and object detection,are customized.Rely on pre-designed models before customizing them.In the image classification model,a backbone network is used to extract the features of the image,and then a fully connected layer is used at the top of the network to output the probabilities on each class.This model achieves up to 99.87%accuracy on the Image Net Te dataset,and its results perform better compared to other models.The Feature Pyramid Aggregation Network(FPANet)is proposed for semantic segmentation based on encoder-decoder structure.In the encoder stage,the model uses a backbone network to extract the features of the image at each resolution,being used as pyramid features,and a lightweight atrous spatial pyramid pooling module is designed at the top of the network for enhancing the semantic information of the features.In the decoder stage,a bilateral directional feature pyramid network is proposed to enhance the extracted features for segmentation,and in this network,a feature pyramid fusion module is proposed to fuse the features on different levels.Finally,a border refinement module is proposed to improve the inaccuracy of the segmentation for small objects and boundary segmentation.The proposed model achieved an accuracy of m Io U of 75.5% and 75.9% on the Cam Vid and Cityscapes datasets,and surpassed the previous model.The One-Stage Detector(OSDet)network is proposed for object detection based on anchor free.It relies on the backbone network to extract the pyramid features,then a feature pyramid network is designed to enhance the representation of the pyramid features,and finally the pyramid features are used for detection.The proposed model achieved an accuracy of up to AP of 82.89% on the publicly available PASCAL VOC dataset,and its results outperformed other models.Second,the model is customized on different computing platforms.In order to make the model perform better,two different network blocks are designed for CPU and GPU respectively in this thesis.Third,the model is customized on different performances.In order to meet the needs of models with the different performances,this thesis designs neural architecture search algorithms to search the backbone networks on which image classification,semantic segmentation and object detection depend,as a way to provide high speed and accuracy models.In addition,the neural architecture search algorithm can also be used to search the network structure of the model under the expected model performance conditions.(2)Design and implementation of model training module.After the user has customized the model,the model will be trained.Since Py Torch itself does not come with a complete training module,this thesis designs a training framework for Py Torch,besides that it also provides data visualization,model evaluation,and other functions.(3)Design and implementation of the model deployment module.After training the model,in order to implement image recognition on the client,the model needs to be deployed to provide an interface for the client to call.This thesis provides two ways to deploy the model,one is to deploy the model on the server,then the user can send a request based on the specification of HTTP interface to complete the image recognition task;the other is to choose to export the model and deploy it on a device such as mobile.(4)Design and implementation of customized image recognition system.This thesis integrate the above image recognition model,model training module and model deployment module,and design and implement a customized image recognition system with web interface.The users can use the system to customize image recognition models without writing code,thus reducing the complexity of customizing models for users.