Cycle-Spinning (CS) method usage in deep learning. Papers: ( Cycle-Spinning Convolution for Object Detection ) and ( Cycle-Spinning GAN for Raindrop Removal from Images )
- Exploits the shift variance of the underlying method.
- Has been successfully used in image de-noising, enhancement, and super-resolution problems.
- Works by applying predefined spatial translations to the input image to obtain several different estimates of the output.
- Gets the final output using the results of aligning and merging these different solutions.
We repeated the experiment by adding Gaussian noise to the Lena image seen in Figure. The results of the convolution, CS applied and low-pass filtered versions of the Lena image are show sequentially. In the upper row of figure, 8 times magnification of Lena’s hat section is shown for better view of the details. As a result, it is seen that the CS method implementation removes noise without losing details. When the results have been compared with the low-pass filtered version, it is also seen that the edge lines of the objects are not blurred in the CS-implemented version.
