CNN Transfer Learning of Shrimp Detection for Underwater Vision System

In deep learning, convolutional neural network (CNN) mostly apply common overland images instead of underwater images classifiers. Even though there are few classifiers that have been introduced in marine and aquaculture application, there is still limited sources of the underwater images such as shrimp images. Generally, most conventional management systems in shrimp aquaculture implemented manual techniques that highly depend on human to observe shrimp conditions. One of the major problems of shrimp aquaculture is the challenge of recognizing underwater images, despite the characteristic atmosphere such as the murky and turbid water conditions. Many models of image classification have been introduced in order to solve the issue of early detection in shrimp and ponds problems. However, there are several limitations of the proposed methods such as semi-intelligence or fully wired systems. Therefore, an intelligence computational method and wireless system or internet of things (IoT)-based system is crucial in making sure a precision aquaculture farming. This study conducted a transfer learning model for CNN real time shrimp recognition. This study aims to accurately assess the performance of the developed CNN model by evaluating shrimp images based on intersection over union (IoU) between annotation and proposed models. The result shows the proposed model can successfully detect the shrimps with more than 95% accuracy. As a conclusion, the proposed model is able to detect the real time video recognition of underwater shrimp in ponds and is applicable in wireless farming.