Research On Data Augmentation Method Based On Data Patterns Supplement

Computer hardware technology has undergone continuous development,greatly alleviating the problem of computing power.This progress has led to a surge in deep learning research.Deep learning has achieved remarkable results in various fields,such as improving image recognition accuracy,making natural language processing more precise and efficient,and enabling object detection at a previously unimaginable level.Obtaining and curating data can be time-,labor-,and resource-intensive,which often presents a bottleneck to improving model performance.The proposal of data augmentation has effectively alleviated this problem.It uses an almost cost-free strategy to sample new data samples from the neighborhood of the current data distribution to expand the current data.While supplementing the amount of data,it also introduces a large number of data patterns,expanding the model’s recognition boundaries.Data augmentation methods are often very simple,such as Translation and Cutout,and there are also some more complex data augmentation methods,such asMixup and CutMix.These data augmentation methods have a common feature-neural networks can see more ”data patterns”.At the same time,we note that many studies have shown that convolutional neural networks can use local information in images to recognize objects,and even very small blocks of information can maintain a certain accuracy.In the field of self-supervision,there are also many methods that make full use of local information to greatly improve the model’s generalization performance.In summary,allowing the model to see more data patterns and making full use of local information can improve the model’s robustness.Inspired by the traditional data augmentation techniques,thesis takes a novel perspective of data patterns and proposes a new method,CyclicShift,to supplement data features and enhance network generalization.Specifically,CyclicShift divides an image into two parts,enabling the model to learn local information for accurate classification.Extensive experiments are conducted on multiple datasets,including CIFAR-10/100,ImageNet,Tiny-ImageNet,CUB-200-2011,Stanford Dogs,Stanford Cars,and Pascal VOC.The results demonstrate that CyclicShift can be applied to various neural network architectures,such as ResNet,PyramidNet,and Wide ResNet,and outperforms existing data augmentation methods.Moreover,we design a new variant method called CyclicMix,which combines the advantages of existing methods and achieves better results than CutMix.To further explain the mechanism behind CyclicShift’s effectiveness,we conduct a series of experiments and visualization analyses.Overall,our study demonstrates the importance of considering data patterns in designing data augmentation techniques and provides a new perspective on enhancing network generalization.