COUNTING TILAPIA FINGERLINGS USING MASK R-CNN

Abstract

This research develops a Mask R-CNN model with a ResNet-101 backbone to count tilapia fingerlings in overlapping conditions. The method is designed to address the limitations of manual counting, which is time-consuming and prone to human error in the fisheries industry. The dataset consists of 4,520 images of 2-3 cm tilapia fish fry taken in a fish farming environment. The process includes data collection, preprocessing, labelling, augmentation, and dataset division. The model is trained using various epochs, with the best performance achieved at the 300th epoch, resulting in a mean Average Precision (mAP) of 0.916 on the training data and 0.819 on the validation data. The evaluation demonstrates the model's ability to count tilapia fish fry by utilising bounding box analysis and Intersection over Union (IoU) to distinguish overlapping fish fry. The research findings show the potential of using computer vision technology to improve tilapia fish farming management and support sustainable aquaculture development.

Access to Document

Other files and links

• Link to publication in Scopus

Fingerprint