A Multi-scale Wavelet CNN for Scanning Electron Microscopy Nerve Image Super Resolution

Efficient acquisition of high-resolution SEM nerve images is an important part of brain science research. Because of the characteristics of SEM nerve image, the SR images obtained by current methods are too smooth and lack detailed information. We first analyze the multi-scale wavelet coefficients of SEM nerve images and find that when decomposed to the third scale, there is almost no structural information at high frequencies and the horizontal components are more obvious than the vertical ones. Based on these characteristics, an end-to-end full convolution neural network based on multi-scale wavelet is proposed. Firstly, the main structure of the network is constructed, which is divided into two modules: encoding and decoding. Then, the actual output of the network is changed to the prediction of multi-scale wavelet coefficients, and the final SR image is obtained by inverse transformation. Finally, multi-objective form is used to set hyper-parameters based on the wavelet characteristics of SEM nerve image. In terms of loss function, two-stage losses (gray-level loss and multi-scale wavelet coefficients loss) are used to emphasize the overall texture and the high-frequency details respectively. In addition, the inverse transformation can replace the deconvolution to reduce some network parameters. In the experiment, the relationship between SR effect and wavelet scale is analyzed. With the increase of wavelet scale, the performance of the model is not linearly enhanced, and the optimal wavelet scale and upscaling factor of the model are determined. Experiments show that, compared with other state-of-the-art methods, our method is more prominent and more realistic in the restoration of nerve structure texture.