224 Automatic image feature extraction for genetic analysis in cattle
Abstract Image analysis has increasingly become an important tool for increasing productivity in many industries, yet its application in breeding programs is under utilized. With coat color patterns from dairy bull images, we explore automatic image analysis that extracts features which can be used in genetic analysis. In order to remove the unnecessary background information, the current methods require time consuming human inspection. Here, we present and compare a composite method that creates a mask (i.e., removes the background portion of the image) and calculates the proportion of dark and light coloration in bulls (n = 657) from the breeds Holstein and Ayrshire in dynamic backgrounds (e.g., forest, grass, hay, snow, etc.). This composite method combines the supervised algorithm MASK-RCNN, an unsupervised image segmentation approach, and k-means color clustering. The first step identifies the region of interest removing the majority of the background noise, while the second and third steps optimize the identification of the bull and segments the color patterning. We find a very low discrepancy between the proportion of white and dark between the manual curation and the composite method (+/- 1.40%); with an immense reduction in data collection time. This automatic composite method greatly improves the efficiency of complex image segmentation and analysis without compromising the quality of the data extracted, making analysis computationally feasible for large data sets. The next step is to calculate genetic parameters from these extracted phenotypes with genomic and/or pedigree data.