Carp-DCAE: Deep convolutional autoencoder for carp fish classification

Abstract

The fisheries industry relies heavily on automatic fish species identification for its socio-economic well-being. Due to the similarity in shape and size of the major carps , it can be difficult to recognise them using morphological features. To recognise these species automatically, our proposed autoencoder network models have been applied to a fish dataset containing 1500 images of three major carps of India. As a feature, the autoencoder model’s latent representation is used. After the training phase is complete, the decoder is removed and fish species are categorised using several classifiers. Different variations of autoencoders, such as the Simple Autoencoder, the Deep Autoencoder, and the Deep Convolutional Autoencoder are applied with different hyper parameters. An encouraging maximum accuracy rate of 97.33% is obtained in 250 epochs with a learning rate of 0.0001 using Deep Convolutional Autoencoder. Some well-known machine learning classifiers, such as Logistic Regression, Naive Bayes, K-Nearest Neighbor, Support Vector Machine, and Random Forest, are also used to evaluate the latent representation’s effectiveness using the latent representation as a feature vector. The Support Vector Machine-based latent representation of the Deep Convolutional Autoencoder outperformed all other approaches significantly. The models’ performance is compared to that of Hu moments, Haralick texture, Weber local descriptor, HOG descriptor etc. with best classifiers along with different deep learning models, such as InceptionV3, InceptionResNetV2, MobileNet, VGG16 and VGG19. The Deep Convolutional Autoencoder outperforms the other models by 52%, 43.55%, 13.77%, 6.67%, 22.22%, 15.11%, 6.66%, 4.89%, and 9.78% respectively. It demonstrates the efficacy of this systematic study in identifying major carps.