| With the rapid development of computer vision and deep learning technology,image classification technology has been widely used in several fields,such as face recognition and object recognition in security monitoring field,lesion recognition in medical imaging field,and traffic sign recognition in autonomous driving field.Although image classification techniques are widely used,they still face many problems and challenges.To address these problems and challenges,the main work and innovations of this thesis can be divided into the following parts:(1)It is an insurmountable obstacle for beginners and non-technical people to use image classification technology,which requires certain knowledge of mathematics,computer science and statistics,using specific programming languages and frameworks.To solve this challenge,this paper designs a zero-code platform for image classification,which is based on the current popular microservice architecture and divides the system into four modules:dataset management,multi-person collaboration,online annotation,and model management according to business requirements,with each module being independent of the other,realizing full decoupling between each module of the system.Through this platform,users can upload datasets,tag datasets,create models,train models,verify models and deploy models without coding and mastering related theories,and develop image classification applications to meet individual needs through graphical operations.(2)Image classification development often requires large amounts of data,but acquiring large amounts of labeled data is costly,and complex models or insufficient training datasets can easily lead to overfitting.To address this problem,the platform has multiple built-in pre-trained models that have been trained on large-scale datasets and have good generalization capabilities that can be used for migration to new tasks.Specifically,the platform uses migration learning techniques to initialize the feature extraction network using pre-trained model parameters,and the user only needs to set up the classification network part so as to meet the needs of image classification applications.In the fine-tuning stage,for cases where the model is too complex or the training dataset is too small,the feature extraction network can be frozen and only the classification layer is optimized;for cases where the user data is not very similar to the pretrained dataset or the training dataset is large,the entire model can be optionally fine-tuned and all model parameters,including the feature extraction network,can be optimized to improve the feature extraction capability of the model under the new task.(3)Image classification models are difficult to deploy and run on edge devices because they require large computational resources and memory for model training and inference.To address this problem,the platform uses a model pruning method to compress the model,which assigns a scaling factor γ to each channel in the convolutional neural network and multiplies it with the channel output,then applies a simple L1 sparsity regularization to these scaling factors.During optimal training,the algorithm trains the neural network weights together with these assigned scaling factors while applying sparsity regularization to the scaling factors to promote channel sparsity.Eventually,for channels with scaling factors converging to zero,the platform crops them according to the user-set compression rate,reducing the computational resources and memory required by the image classification model,thus making it suitable for edge devices. |
