Multi-species Fish Identification using Hybrid DeepCNN with Refined Squeeze and Excitation Architecture

Abstract

Fish play a prominent role in the food web and fish farming has value for both human consumption and tourist attractions. Due to the increasing importance of marine biodiversity, recognition of fish species has become a prominent task in monitoring the mislabelling of seafood and extinct species. This problem can be solved using traditional manual annotation on the images. To reduce manpower, cost, and tremendous time, deep learning approaches are used which always require large datasets. Therefore, fish species identification is a challenging task using disproportionately small data sets. In this research, we develop a new method by refining the squeeze and excitation network for the automatic fish species classification model to identify 23 different types of fish species. To achieve this, a hybrid framework using deep learning is proposed on a large-scale dataset and implemented transfer learning for a small-scale dataset. Deep learning methods can be used to identify fish in underwater images. In this study, we have proposed a new method of hybrid Deep Convolutional Neural Network (CNN) along with a Support Vector Machine (SVM) for classification. Additionally, the Squeeze and Excitation (SE) block has been improved for improved feature extraction. The proposed method achieved an accuracy of 97.90%. Then post-training with the small-scale dataset (Croatian) achieved an accuracy of 94.99% with an 11% improvement compared to Bilinear CNN (B-CNN) (Qui et al., 2018) and can be used in any underwater applications to identify fish species and avoid mislabelling of seafood.