This talk uses image processing as an example to explore the importance of geometrically equivariant priors in deep network design. It focuses on introducing the construction methods and basic theories of novel network modules such as high-precision rotation/scale equivariant convolution and rotation-equivariant implicit neural representation. Furthermore, this talk will demonstrate through practical applications such as medical/natural image processing and image reconstruction that embedding prior geometric symmetry can significantly improve the performance and generalization ability of the model.