Jisuanji kexue yu tansuo (Nov 2024)
Image Data Augmentation Method for Random Channel Perturbation
Abstract
The simulation of object occlusion strategies in data augmentation sets all the pixels in the randomly cropped region of the input image to zero, which erases the effective texture features and leads to poor network generalization. Therefore, this paper proposes a novel data augmentation method known as the “ChannelCut” method. The “ChannelCut” includes two methods: ChannelCut1 and ChannelCut2. Firstly, three square regions are randomly selected on the input image, and the channels of the input image are split to three channel images. Secondly, the ChannelCut1 method selects a square region on the three channel images respectively. The pixels selected by the three channels are different from each other and are set to zero. At the same time, the ChannelCut2 method retains the pixels of the square area selected on each channel in the ChannelCut1 method, and the pixels of the other two square areas corresponding to the channel are set to zero. Finally, the two methods merge the three channel images together to obtain two random channel perturbed images. The proposed method is fused into CNN models such as Resnet18, ShuffleNet V2, MobileNet V3 and experiments are carried out on five datasets such as CIFAR-10 and Image-nette. The results show that the proposed method has a better classification accuracy than the mainstream method on five datasets. Furthermore, the baseline performance has shown a significant improvement. The proposed method has advantages in fine-grained image classification and outperforms the automatic data enhancement type method that uses reinforcement learning in terms of time performance. The ChannelCut method has strong generality and effectiveness, can retain image texture features to different degrees, and enrich image diversity, significantly improving the robustness and generalization of the convolutional neural network model.
Keywords
