FISH AGE ESTIMATION FROM OTOLITH CONTOURS AND FISH SCALES USING CONVOLUTIONAL NEURAL NETWORK

• Main Author: Lailani, Ade
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/71891
• Institution: Institut Teknologi Bandung

Estimating the age of fish is one of the keys to studying fish populations, providing useful knowledge about growth rates, mortality, and maturity ages. For many fish species, hard structures such as otoliths and fish scales can be analyzed to estimate the age of the fish. Generally, expert fish age reading methods are based on reading fish scales and otolith images, which is often a time-consuming and expensive process that requires an automated and cost-effective approach. Convolutional Neural Networks (CNNs) is a deep learning class for analyzing images. The use of the CNN method requires modification of its architecture to make the network more efficient and provide up-to-date accuracy in the processing provided. The CNN models built are Simple CNN, Stacked CNN, and Deeper CNN models. In addition, the implementation of EfficientNet transfer learning was carried out. Image processing and testing of the CNN model are carried out using the GPU from Google Colab Pro. The simple CNN model provides greater accuracy than the other two models, namely stacked CNN and deeper CNN. The average accuracy of simple CNN is 96% followed by stacked CNN with an average accuracy value of 95% and deeper CNN of 93%. The deeper CNN model has not been able to achieve a higher accuracy value, but the use of the deeper CNN model has the fastest time of the other two models, namely 59 minutes. Adam's optimizer is better than adadelta to get high accuracy values. Meanwhile, the Adadelta optimizer is better than the Adam optimizer in time efficiency. The use of efficientnet-B0 to B4 gives the same accuracy and slightly different loss values. The difference is in the time when the efficientnet-B0 has the fastest time compared to other efficientnet models. While efficientnet-B1, efficientnet-B2, efficientnet-B3, and efficientnet-B4 respectively have a longer time. The CNN+Efficintnet-B0 model has the best performance in terms of speed, reaching 8 minutes, which means 7x faster than the CNN model without efficientnet.