Fish Image Classification Based on Feature Extraction Using Convolutional Autoencoder and CNN Classifier

Abstract

Automatic fish species classification is crucial for supporting fisheries management, particularly in coastal regions such as Lombok. This study proposes an artificial intelligence-based approach combining a Convolutional Autoencoder (CAE) as a feature extractor and a Convolutional Neural Network (CNN) as a classifier. The dataset consisted of 9,000 colored images from nine fish species, which were converted to grayscale. Experimental results show that the CAE+CNN model achieved a classification accuracy of 95.94% within 0.87 seconds for 1,800 images. In comparison, the pure CNN model achieved a higher accuracy of 99.05% and faster performance, but with a much larger model size and complexity. The CAE+CNN approach is considered more lightweight and efficient, making it suitable for deployment on resource-constrained devices. This technology has strong potential for local implementation to support the digital transformation of the fisheries sector and assist fishing communities in automatically identifying fish species, improving efficiency, and reducing dependency on manual classification by experts.