IES: A Powerful Visual Feature Representation Network.

Effective visual information representation is significant for the feature extractability of deep neural networks. The transformation from images to feature maps realized by the stem, referring to the initial part of the neural network that processes input images, causes the loss of information owing to the heterogeneity of the color and structure information: the most prominent features. To solve this problem, we propose a powerful and effective image-embedding stem (IES) model with a color-embedding module (CEM), structure embedding module (SEM), and feature mixing module (FMM). Specifically, the red-green-blue (RGB) ternary color information is embedded into a high-dimensional vector space containing rich feature information through the CEM. Simultaneously, the SEM is used to explicitly encode multiscale structural information to enrich the detailed information in the feature maps. Finally, they are fused by the FMM to preserve more details. Comprehensive experiments demonstrate the efficacy of the IES in different visual tasks. It achieved +1.2 and +0.5 top-1 accuracy ratings on the ImageNet-100 dataset for the VanillaNet-5 and TinyViT-5m backbones, respectively, and obtained +2.36 and +1.7 mean intersection-over-union scores on the UTFPR-SBD3 dataset for PoolFormer and ConvNeXtV2 backbones, respectively. The code and models will be released soon. [ABSTRACT FROM AUTHOR]