Wrinkle, Lighting & Shadow Adjustment
Deep Learning 0
Wrinkle removal is an important part of any image processing workflow which involves garment editing/retouching. Although current image editing softwares provides the tools for removing wrinkles from a garment image, it requires expert image editing skills to use these tools for wrinkle removal.
Broadly, wrinkles are formed in garment photographs due to two major reasons –
- Highlight and Shadow: The gradient from dark regions (shadow) to lighter regions (highlight) gives the perception of a wrinkle.
- Folds: In many cases, garment wrinkles are formed due to folds. These are more prominent in pattern garments because the pattern gets distorted due to displacement.
Some popular techniques for manual wrinkle removal include –
- Dodge and Burn – Dodge (lighten) and burn (darken) tools are used to adjust the lighting near the wrinkles in an image such that the wrinkles are no longer visible. Similar effects can also be achieved using curves adjustment and appropriate masking. [Reference]
- Frequency Separation – Separate the input image into its high frequency (texture) and low frequency (color) components using blur and subtraction. Fine fabric textures are captured in the high frequency component, and wrinkles are captured in the low frequency component. Finally, the wrinkles are softened in the low frequency channel using the mixer brush tool. [Reference]
However, these techniques require human intervention at multiple stages –
- Domain Expertise is required to identify wrinkles in a garment image. This is not trivial because wrinkles could be easily confused with folds from a skirt or a women’s dress. There is a vast variety of garment images (visual features such as color, texture and shape), which makes the process of identifying wrinkles more difficult.
- Technical Expertise: Once all the wrinkles have been identified, a human expert must process each wrinkle individually using techniques described above. It is not trivial to automate this process, because the choice of the tool depends on the type of the wrinkle. There is also a need for fine-tuning the settings for each tool depending on the intensity of the wrinkle and the desired wrinkle removal effect, making the process tedious.
At FlixStock, we employ deep learning techniques to overcome these challenges. A high-level overview of the approach is described below –
- Garment Mask Segmentation – A fine-grained garment segmentation mask is predicted for the input image.
- Wrinkle Segmentation – A fine-grained wrinkle segmentation mask is predicted for the detected wrinkles in the input image.
- Highlight and Shadow Adjustment – High resolution highlight and shadow adjustment is carried out to remove wrinkles formed due to lighting inconsistencies.
- Pattern Retouching – In case of pattern distortion, patterns are regenerated in the detected wrinkle region.
Sample results of our approach are shown below.