| Image classification is a fundamental task in computer vision,which is widely adopted in real-world applications.The current methods for image classification are mostly based on deep learning,especially deep convolutional neural network(DCNN).Despite of the massive improvement compared to traditional methods,the image clas-sification based on DCNN still have many drawbacks and the performance is not sat-isfactory for real-world applications.This dissertation focuses on four sub-tasks in im-age classification,i.e.,generic image classification,fine-grained image classification,multi-task incremental learning and multi-class incremental learning,to gradually deal with the practical challenges for this task.Specifically,generic image classification is the fundamental task in image classification,fine-grained image classification needs to handle subtle inter-class difference,and incremental learning aims to solve the prob-lems brought by the continuous streams of incoming data.The main contributions are summarized as follows:(1)A learning framework named DualNet based on complementary feature learn-ing,which is designed for generic image classification.The extraction of visual features is usually treated as the most important design for image classification,while a single DCNN cannot learn all the details of input images.DualNet coordi-nates two sub-networks to learn features complementary to each other by adding constraints in the training process,and thus the features after fusion can be more accurate.The extensive experiments show that DualNet can achieve higher clas-sification accuracy than single network or model ensemble.(2)A learning framework named HybridNet based on multi-granularity feature learn-ing,which is designed for fine-grained image classification.The fine-grained image classification needs to handle more subtle inter-class difference and larger intra-class variation,which is more challenging than generic image classifica-tion.HybridNet is motivated to fully exploit the label hierarchy of fine-grained images as supervision to learn multi-granularity features,which are then fused to form more discriminative representation of the input images.The experiments on multiple fine-grained datasets demonstrate that HybridNet can achieve higher classification accuracy than the baselines such as Compact Bilinear CNN.(3)A novel method for multi-task incremental learning,which is composed of Dis-tillation and Retrospection.In practice,the data of different tasks usually comes continually due to privacy or memory cost.In the proposed method,the target model adapts to the new task by knowledge distillation from an intermediate ex-pert,while the previous knowledge is more effectively preserved by caching a small subset of data for old tasks.The extensive experiments on multiple task sequences demonstrate that the proposed approach can bring consistent improve-ments on both old and new tasks.(4)A novel method for multi-class incremental learning,which is composed of Co-sine Normalization,Less-forget Constraint and Inter-class Separation.Multi-class incremental learning aims to learn a unified classifier on all the observed classes in each phase,which is more challenging than multi-task incremental learning.Our study reveals that the imbalance is a crucial factor affecting the per-formance of multi-class incremental learning,and the proposed method mitigates the adverse effects of the imbalance from different aspects such as network nor-malization and class boundary.Experiments on CIFAR100 and ImageNet show that the proposed method can outperform the existing methods by a large margin.In summary,this dissertation dives deeply into the task of image classification.According to the characteristics of real-world applications,several methods are pro-posed in corresponding to the problems of the existing works for image classification.The extensive experiments demonstrate that the proposed methods can bring significant improvements,and are greatly valuable for real-world applications. |
PDF Full Text Request