Research On Lifelong Learning Classification Model Based On Image Similarity

At present,various deep learning models have become mature in the field of computer vision,such as image classification,object detection,semantic segmentation and other tasks,these models emerge in an endless stream with high accuracy.However,in so many models,the problem of catastrophic forgetting is still unavoidable.Catastrophic amnesia refers to the fact that a model can only adapt to one task,and when the next task is performed,the knowledge of the previous task is largely forgotten,resulting in a significant decrease in the accuracy of the model on the previous task.To solve this problem,researchers have proposed the concept of lifelong learning.The core idea of lifelong learning is to learn new tasks while retaining previously learned knowledge,and make the previous tasks have a positive impact on the learning of new tasks,rather than a negative impact.Lifelong learning allows machine learning systems to constantly adapt to new environments and tasks without having to retrain models.Based on the current research status and the problem of catastrophic forgetting,this thesis proposes a lifelong learning model based on image similarity.The model mimics human memory,in which when people recognize an image,they distinguish it by comparing it to the memory.In order to realize this model,the main research work of this thesis includes the following three aspects: First,the existing imagenet data set is divided into multiple subtasks,and the appropriate category template is selected to build multiple data sets,so as to facilitate the learning and evaluation of the model.Secondly,a feature extraction network of multi-level feature combination is established to extract and fuse the template data and image data respectively.The accuracy of the feature extraction model can reach more than 90%when the classification task experiment is carried out alone,which proves that the network can extract enough features of the image.Then,according to the existing image similarity comparison method and deep learning model,a classifier is established to compare the features extracted from the image,so as to realize the task of image classification.Finally,after the completion of one task,the next task was trained to evaluate the results of the new task.At the same time,the accuracy of the first task was directly evaluated,and the accuracy of each task was recorded.The general evaluation method of lifelong learning was used to evaluate the advantages based on image similarity.After the actual experiments and the verification of relevant theories,the structure of the model is improved to make the model more accurate in a single task.At the same time,lifelong learning with sustainable learning ability is established,and compared with the general neural network,showing the advantages of lifelong machine learning.