New Label Learning For Multi-Label Image Classification

Multi-label learning is an important learning paradigm in the field of machine learning.In the past few years,multi-label learning has been widely applied to many learning scenarios.With the continuous development of multi-label learning,the performance of multi-label learning algorithms is becoming more and more mature.At the same time,dynamic environments pose new challenges to the machine learning field on the adaptability of algorithm.Previous multi-label learning algorithms assum that the label space is consistent in prediction phase and training phase,while this assumption may not hold in real dynamic environments.Consequently,it is necessary to proposes adaptive algorithms for multi-label learning in dynamic environments.Multi-label images are a significant kind of multi-label learning objects.Multi-label image classification task is currently challenging research topic,and it has broad application prospects in the video field of the Internet and security surveillance scenarios.In recent years,the development of convolutional neural network(CNN)technology has greatly accelerated the development of image recognition,meanwhile the model pre-trained on large-scale training sets can be migrated to other related learning tasks.Therefore,it is of great theoretical value and application value to combine the convolutional neural network technology to study the new category label learning problems that may be encountered in multi-label image classification tasks.This thesis will focus on the multi-label image classification task in the dynamic open environment,and design an algorithm to solve the problem of the new category label.In this thesis,the new class label learning algorithm MULNLC is proposed.Firstly,the pre-trained and fine-tuned convolutional neural network model are used as the feature extractor of multi-label images,and then classification model and new label detection model are learned based on the initial training set.After that the detetion model is applied to the batch data involving new label to detect new label instances,and then the instances identified as new label instances are further expanded by the convolution network model.Finally,the classification model and detection model are updated based on new class label data and initial known label data.The experimental results on the multi-label image benchmark datasets show the effectiveness of the algorithm.Meanwhile,based on the proposed MULNLC algorithm,an image anotation system for new class label detection is developed for practical application scenarios.