https://medium.com/@amanrao032/image-upscaling-using-bicubic-interpolation-ddb37295df0
Other forms of basic upscaling
- Nearest-Neighbor (NN)
- Copies the value of the closest pixel without any blending.
- fast but tends to produce blocky and pixelated results
- Bilinear Interpolation
- Averages the values of the 4 closest pixels to estimate a new pixel value.
- smoother results compared to NN but may still lack fine details.
- Lanczos Resampling
- Uses a mathematical sinc function over a larger area (e.g., a 3×3 or 4×4 grid).
- preserves sharpness and details but can be computationally intensive.
Bicubic interpolation takes into account 16 surrounding pixels (a 4x4 grid), which is more than what Nearest Neighbor or Bilinear Interpolation uses. This approach results in smoother transitions and better preservation of image details.
Sample Implementations code
- https://github.com/cdsvitbhopal/ProjectArena-ML/blob/main/Image Upscaling using Bicubic Interpolation/bicubic.py
- Explanation of the code
Handling edge cases Corner and edge pixels do not have a 4x4 neighbourhood so interpolated their values using the nearest neighbour method which involves setting their value to the that of the nearest known pixel but other techniques like clamping, mirroring and wrapping can be used to interpolate their values.
some examples done
