A New LCFCN-based Approach for Weakly-Supervised Fish Segmentation

Abstract

Fish statistics and measurements are important for aqua-environment nowadays. While physical approaches might be expensive and prone to be erroneous, some automatic methods are dependent on the necessity of full annotations for supervised segmentation process; which is time-consuming and required manual labors. Inspired by the deep-learning methods and weakly-supervised approaches, we develop an efficient network for dataset with point-level supervision that fishes are labeled in a single mouse-click. With one branch containing our proposed baseline network output, the other branch includes the affinity matrix output that both of them are concatenated before being fed into a random walk architecture to attain the final. Additionally, the proposed architecture is trained with a new loss function based on the localization-based counting fully convolutional neural network (LCFCN); before being validated on the FishSeg testing part of the DeepFish dataset. Experimental results have confirmed the validity of our proposed affinity-LCFCN (A-LCFCN) solution on such a cheap fish dataset conbining both point-labeled and fully-masked images.