Contour Extraction of Individual Cattle from an Image Using Enhanced Mask R-CNN Instance Segmentation Method

Abstract

In animal husbandry, the traceability of individual cattle, their health information, and performance records greatly depend on computer vision and image processing-based approaches. However, some of these approaches perform below expectations in obtaining real-time information about individual cattle. No doubt, inaccurate segmentation and incomplete extraction of each cattle object from an image are notable contributory factors. As accurate segmentation is a prerequisite for obtaining real-time information about individual cattle, and since the algorithm of Mask R-CNN relies on the algorithm of simultaneous localization and mapping (SLAM), for the construction of the semantic map, which sometimes exchanges image background for the foreground, there is a need to enhance the available approaches towards achieving precision animal husbandry. To achieve this, an enhanced Mask R-CNN instance segmentation method is proposed to support indistinct boundaries and irregular shapes of cattle bodies. The methods employed in the research are in multiple folds: (1) Pre-enhancement of the image using generalized color Fourier descriptors (GCFD); (2) Provision of optimal filter size that was smaller than ResNet101 (the backbone of Mask R-CNN) for the extraction of smaller and composite features; (3) Utilization of multiscale semantic features using region proposals; (4) A fully connected layer of existing Mask R-CNN integrated with a sub-network for enhanced segmentation and (5) Post-enhancement of the image using Grabcut. Experiments on the datasets of cattle images produced better results when compared to other state-of-the-art methods with 0.93 mAP.