Conv2NeXt: Reconsidering Conv NeXt Network Design for Image Recognition

Currently, transformer in computer vision, whether in image recognition, image segmentation or target detection, are widely popular due to their excellent performance. Traditional convolutional neural networks (CNNs) are dwarfed by Transformers, but the disadvantage of Transformers is that they require huge resources for cultivation. In 2022, Liu et al. rethought CNN s and proposed a series of models that refer to the structure of Transformers (e.g., Swin Transformers). In contrast to Transformers, Liu's CNN models, dubbed ConvNeXt, use the standard convolutional neural network. The accuracy of ConvNeXt on the ImageNet dataset is surprisingly close to that of Swin Transformers, proving that CNN s still have superior performance. However, ConvNeXt does not perform as well on low-resolution datasets, such as CIFAR-10, CIFAR-100, and Tiny ImageN et, and there is still much room for improvement. To enhance the generalisation of ConvNeXt, we propose Conv2N eXt models, which consist of two patch embedding layers and a symmetric convolutional block of repeated applications. In addition, Conv2N eXt incorporates some attention modules to refine feature maps, making them more robust. In experiments, Conv2NeXt surpasses most well-known and widely used state-of-the-art performances (SOT As) with an accuracy of 98.23 % in CIFAR-10, 83.82% in CIFAR-100, and 71.94% in Tiny ImageNet. Most importantly, these performances were trained by only one customer-based GPU, achieving almost SOTAs with a much lower number of parameters (Params). The open source code and checkpoint of this work are accessible at https://github.com/southerly7/Conv2NeXt.