Weakly registered fluorescence image super-resolution method based on residual DualGAN

Fluorescence microscopic images can reflect the fine structure of biological tissues at the micrometer level, which is of great value for morphological studies. Imaging a large range of tissues has problems such as high cost and poor resolution. Image super-resolution methods can improve the image resolution at a low cost. However, image super resolution methods require precisely registered low- and high-resolution image samples, which is very difficult to obtain. We propose an image super-resolution method that can be used for weak registered image samples. This method is based on DualGAN to generate correct super-resolution results, avoiding the requirement of pixel-level registration. The residual network and attention mechanism are incorporated to improve performance of the network. We demonstrated that our method along with other methods enhanced the resolution of the neural fluorescence image dataset by threefold. The results show our method achieves the best super-resolution reconstruction.