Segmentation and graph generation of muzzle images for cattle identification

Abstract

The current methods for the organizing the records (i.e., cataloguing) of cattle are known to be archaic and inefficient, and often harmful to the animal. Such methods include the use of metal tags attached to the animal's ears like earrings and of branding irons on their necks. Previous research on new methods of livestock branding based on computer vision techniques utilized a mixture of texture features such as Gabor Filters and Local Binary Pattern as a means of extracting identifying features for each animal. The presented approach proposes a new technique using the muzzle image as an individual identifier as a novel technique, assuming that the muzzle RoI taken as input for the model pipeline is already extracted and cropped. This task is performed in three steps. First, the muzzle image is segmented via a convolutional neural network, resulting in a bitmap from which a graph structure is extracted in the second phase. The final phase consists of matching the resulting graph with the ones previously extracted and stored in the database for an optimal match. The results for the segmentation quality show a fidelity of around seventy percent, while the extracted graph perfectly represents the extracted bitmap. The matching algorithm is currently in progress.