Research On Several Issues Of Image Generation And Recognition Based On Deep Learning

In the last decade,deep neural networks(DNNs)have been widely used in several real-world applications,such as image generation and pattern recognition.However,these advanced models are still lacking in theoretical analysis and become limited to more complex tasks.Specifically,the generalization in DNNs is not sufficiently studied,and their performance in real-world testing depends on the corresponding generalization capability.Besides,the training of DNNs can be unstable with novel objective functions,especially for generative adversarial networks.Moreover,DNNs generally require a large amount of data for model optimization,limiting the application for small data analysis.Therefore,this thesis introduces novel formulations to modify the existing DNNs,aiming at improving their performances of computer vision tasks,including image inpainting,few-shot learning,and image classification.Main contributions are listed as follows:1)This thesis proposes a progressive generative network(PGN)for image inpainting.Considering it is difficult to repair the natural images with large broken areas at once,PGN formulates semantic image inpainting as a curriculum learning problem.Specifically,PGN first divides the whole inpainting process into several phases,where each phase is designed to finish a course of the entire curriculum.After that,an LSTM model is used to string all the phases together.By introducing this curriculum learning strategy,the generalization capability of DNNs can be significantly improved.PGN can,therefore,progressively shrink the corrupted regions and yield promising inpainting results.As the entire inpainting process is performed in a single forward pass,PGN is much faster than the existing methods.Extensive experiments on Paris Street View and Image Net datasets demonstrate the superiority of our approach.2)This thesis develops a Deep Image Co-Inpainting(DICI)framework.The motivation behind this is that natural images may produce inconsistent reconstruction loss to affect the training process in image inpainting,which reduces image inpainting performance.Inspired by co-distillation,DICI integrates two generators and two discriminators to address the above issue.Specifically,the two generators utilize an image-level co-distillation to output consistent image inpainting results,which reduces the external chaos;the two discriminators introduce a feature-level co-distillation to produce a consistent judgment,which alleviates the catastrophic forgetting in their training processes.The two generators and two discriminators perform cross-network training to improve stability.This thesis tries five different implementations of DICI and compares them with the existing methods on four datasets to verify the superiority.3)This thesis proposes Miner-Evaluator Curriculum Network(MECN)for few-shot learning.Few-shot learning studies how to train DNNs with few samples.The few-shot tasks are randomly sampled,which causes omissions and cannot guarantee their quality.MECN introduces two assistant modules to address the above issue: a)Task Miner samples hard few-shot tasks according to the feedback from the meta-learner,and b)Task Evaluator re-weights the few-shot tasks according to their noise levels.Since the two modules are collaborative and complementary,we integrate them to provide the metalearner with a task-level curriculum and improve its generalization performance.To evaluate MECN,we conduct extensive experiments in both supervised and unsupervised settings of few-shot learning on two challenging datasets,including mini Image Net and tiered Imagenet.We additionally perform ablation studies to verify the high effectiveness and efficiency of our approach.4)This thesis presents Adversarial Co-distillation Network(ACN)for image classification,which improves the co-distillation by generating extra divergent examples.Specifically,ACN includes a generative adversarial module built upon Generative Adversarial Networks(GAN)to generate the divergent examples and a co-distillation module consisting of two classifiers to learn the divergent examples.These two modules are learned in an iterative and adversarial manner.Moreover,we introduce weakly residual connections and restricted adversarial searching to guarantee the quality of the divergent examples and improve the training stability.Extensive experiments with various deep architectures on different datasets demonstrate the effectiveness of ACN.