AI Feature Extraction for Endurance Race Photos

Project

• Endurance Race companies and photographers need cost-effective ways to advertise their services.
  • Social media is proven the best way for them to advertise.
  • They don’t have expertise in social media.
  • Social media experts are too expensive for the industry.
• This project will help create AI content bots that grow their social media presence by automatically posting the best photos from each race.
• The PhotoHive_DSP algorithm extracts features from race photos so they can be fed into a machine learning model that finds the best photos. It measures:
  • Sharpness of people and bikes in the image,
  • Motion blur direction and intensity,
  • Color Palette of the image,
  • Brightness and contrast of color channels, and
  • Image color saturation.

System

Methods

• Motion blur is calculated by finding streaks, then locating where the frequency response orthogonal to streaking crosses a minimum threshold.
• Sharpness is calculated using the variance of the Laplacian of the image.
• Color Palette is extracted using the Adaptive Octree Color Quantization algorithm, taking notes from the Octree Grouping and K-Means Clustering.

Conclusion

• Results
  • The color palette has been subjectively verified by a sample population of 20.
  • The sharpness measure is biased according to the linear scaling of subject size.
  • The vectors of motion blur are reasonably accurate.
• Possible algorithm improvements
  • There is significant error in the Blur Profile due to cartesian-to-polar conversion in corners and lack of a windowing function to prevent the Gibbs’ Phenomenon.
  • The traditional sharpness calculation is biased toward any high frequency data, and small bounding boxes, not just hard edges. Instead, use the LoG filter and swap variance, ${\sigma }^2=\frac{\sum _{𝑛=0}^{\mathrm{N-1}}({{𝑥}_{𝑛}-𝜇}^2)}{N}$ , for ${\sigma }^2=\frac{\sum _{𝑛=0}^{\mathrm{N-1}}{({𝑥}_{𝑛}-𝜇)}^{(3 \; or\; 4)}}{N}$