Fig. 1. Overall structure of multi-module combined denoising network model

Fig. 2. Structure design of channel association module, adopting different structures in the first and last layers, but the same structure in the middle layers

Fig. 3. Structural design of the multi-scale denoising module, including the information extraction layer of different scale convolution kernels, the middle U-Net layer and the upsampling layer of different magnifications

Fig. 4. Structure of fusion compression module

Fig. 5. Examples of scanned areas of animal and plant samples and their corresponding Ground Truth images. (a) Slice of Cyclops; (b) hair follicle slice; (c) lichen slice; (d) stem longitudinal cut of terminal apex

Fig. 6. Changes of some parameters during network model training and recovery performance at different stages of training. (a) Change trend of loss value of network model in the training set and accuracy in the verification set; (b) recovery effect of a randomly selected image at different stages of network model training

Fig. 7. Visual effect of cross-section image data of bluegrass leaves processed by different denoising methods. (a) Noise image; (b) repair result of Gaussian filter; (c) repair result of BM3D; (d) repair result of DnCNN; (e) repair result of FFDNet; (f) repair result of CBDNet; (g) repair result of DNMMC

Fig. 8. Visual effect of image data of fish gill cross-section samples processed by different denoising methods. (a) Noise image; (b) repair result of Gaussian filter; (c) repair result of BM3D; (d) repair result of DnCNN; (e) repair result of FFDNet; (f) repair result of CBDNet; (g) repair result of DNMMC

Fig. 9. Visual effect of image data of spore cup samples processed by different denoising methods. (a) Noise image; (b) repair result of Gaussian filter; (c) repair result of BM3D; (d) repair result of DnCNN; (e) repair result of FFDNet; (f) repair result of CBDNet; (g) repair result of DNMMC