Food image denoising faces dual challenges of preserving complex textures while suppressing noise. To address limitations of traditional methods—including fixed receptive fields struggling to capture multi-scale features and up-sampling processes lacking effective noise suppression causing texture blur—this study proposes ECAMixDNet, a hybrid network integrating Swin-Transformer with CNN decoders. The framework introduces a CAMixer dynamic encoding module that adaptively captures irregular food texture features through self-adjusting mechanisms. For feature reconstruction, we innovatively design an Efficient Channel-Aware Mixed Up-sampling (ECAMixUp) method to achieve multi-scale channel noise suppression, effectively reducing high-frequency artifacts. To overcome data scarcity in food denoising research, we constructed the first paired noise-clean food image dataset, DeNoiseFood. Experiments demonstrate that ECAMixDNet preserves critical texture details and color fidelity on DeNoiseFood, outperforming baseline models in PSNR/SSIM metrics. Cross-domain validation on the SIDD dataset confirms its generalization capability, providing robust technical support for food recognition, nutritional analysis, and related applications. The code for this paper can be found at: https://github.com/YX-yy/ECAMixDNet .