Confocal microscopy is one of the most widely used tools for high-resolution cellular, tissue imaging and industrial inspection. Micrograph reconstruction based on deep learning has become an effective tool for modern microscopy imaging techniques. While most deep learning methods neglect the imaging process mechanism, which requires a lot of work to solve the multi-scale image pairs aliasing problem. We show that these limitations can be mitigated via an image degradation model based on Richards-Wolf vectorial diffraction integral and confocal imaging theory. The low-resolution images required for network training are generated by model degradation from their high-resolution counterparts, thereby eliminating the need for accurate image alignment. The image degradation model ensures the generalization and fidelity of the confocal images. By combining the residual neural network with a lightweight feature attention module with degradation model of confocal microscopy ensures high fidelity and generalization. Experiments on different measured data report that compared with the two deconvolution algorithms, non-negative least squares algorithm and Richardson-Lucy algorithm, the structural similarity index between the network output image and the real image reaches a high level above 0.82, and the peak signal-to-noise ratio can be improved by more than 0.6 dB. It also shows good applicability in different deep learning networks.