An Instance Segmentation Method for Nesting Green Sea Turtle’s Carapace using Mask R-CNN

Abstract

This research presents an improved instance segmentation method using Mask Region-based Convolutional Neural Network (Mask R-CNN) on nesting green sea turtles’ images. The goal is to achieve precise segmentation to produce a dataset fit for future re-identification tasks. Using this method, we can skip the labour-intensive and tedious task of manual segmentation by automatically extracting the carapace as the Region-of-Interest (RoI). The task is non-trivial as the image dataset contains noise, blurry edges, and low contrast between the target object and background. These image defects are due to several factors, including jittering footage due to camera motion, the nesting event occurring during a low-light environment, and the inherent limitation of the Complementary Metal-Oxide-Semiconductor (CMOS) sensor used in the camera during our data collection. The CMOS sensor produces a high level of noise, which can manifest as random variations in pixel brightness or colour, especially in low-light conditions. These factors contribute to the degradation of image quality, causing difficulties when performing RoI segmentation of the carapaces. To address these challenges, this research proposes including Contrast-Limited Adaptive Histogram Equalization (CLAHE) as the data pre-processing step to train the model. CLAHE enhances contrast and increases differentiation between the carapace structure and the background elements. Our research findings demonstrate the effectiveness of Mask R-CNN when combined with CLAHE as the data pre-processing step. With CLAHE technique, there is an average increase of 1.55% in Intersection over Union (IoU) value compared to using Mask R-CNN alone. The optimal configuration managed an IoU value of 93.35%.