Regency C
Statistical Image Processing for Machine Learning (303984)
*Vikram Krishnamurthy, Alliance Innovation Lab Silicon Valley
Kusha Nezafati, Alliance Innovation Lab Silicon Valley
Keywords: machine learning, spatial bootstrapping, statistical distance, image analytics
Image classification utilizing Machine Learning requires a large amount of training data. For applications in predictive maintenance (e.g. component wear, equipment failure, etc.), training data is limited due to prohibitively high cost and extended time frames required for collecting data. In this paper, we discuss the use case for air filter wear where a statistical distance metric is: (1) used as a correlated attribute to wear, and (2) computed using pixel intensity distributions of images of air filters at specific wear levels compared to a clear air filter. The statistical distance metric is utilized as an attribute for Machine Learning to predict air filter wear and lifetime. Due to limited training data, a sampling distribution of the statistical distance metric was derived utilizing spatial bootstrapping. As these images exhibited high spatial pixel to pixel correlation, a block bootstrap approach was used where the block length was selected based on variogram analysis and overlap sampling variance estimators. This approach limited the likelihood of overfitting due limited training data. A bias-variance tradeoff analysis was conducted of the key spatial block bootstrapping