Evaluation of ResNet network for semantic segmentation of coral reefs

Abstract

The Deep Residual Network (ResNet) learning model is known to achieve better accuracy and requiring shorter training time compared to other pre-trained learning models for image classifications and recognition. In this paper, the use of ResNet networks for semantic segmentation of coral reefs images was explored. Three ResNet networks (ResNet-18, ResNet-50, and ResNet-101) were evaluated and compared using 900 images as training dataset and 38 images as test dataset. The last three layers of the pre-trained ResNets were replaced with a set of layers that classified each pixel of the images into four classes: ‘dead’, ‘alive’, ‘sand’ and ‘unknown’. A Softmax layer was introduced to reduce the imbalanced defects. Then, DeepLabv3+ employed the Atrous convolution to extract the features computed by applied CNN and segment the pixels of the object. ResNet-101 was shown to achieve better results compared to ResNet-18 and ResNet-50. Further analysis of the results implied that the class weightage assignment needs to be improved and more a larger training dataset should be acquired.