FishResNet: Automatic Fish Classification Approach in Underwater Scenario

Abstract

Fish species classification in underwater images is an emerging research area for scientists and researchers in the field of image processing. Fish species classification in underwater images is an important task for fish survey i.e. to audit ecological balance, monitoring fish population and preserving endangered species. But the phenomenon of light scattering and absorption in ocean water leads to hazy, dull and low contrast images making fish classification a tedious and tough task. Convolutional Neural Networks (CNNs) can be the solution for fish species classification problem but the scarcity of ample fish images leads to the serious issue of training a neural network from scratch. To overcome the issue of limited dataset the present paper proposes a transfer learning based fish species classification method for underwater images. ResNet-50 network has been used for transfer learning as it reduces the vanishing gradient problem to minimum by using residual blocks and thus improving the accuracies. Training only last few layers of ResNet-50 network with transfer learning increases the classification accuracy despite of scarce dataset. The proposed method has been tested on two datasets comprising of 27, 370 (i.e. large dataset) and 600 images (i.e. small dataset) without any data augmentation. Experimental results depict that the proposed network achieves a validation accuracy of 98.44% for large dataset and 84.92% for smaller dataset. With the performance analysis, it is observed that this transfer learning based approach led to better results by providing high precision, recall and F1score values of 0.94, 0.85 and 0.89, respectively.

Targeted Data Augmentation and Hierarchical Classification with Deep Learning for Fish Species Identification in Underwater Images

“…The output of the first fully connected layer is of 9216 units instead of 4096 units as in the original AlexNet. Mathur et al [ 31 ] fine-tuned the ResNet50 model by retraining only the last fully connected layers without any data augmentation. They used Adamx as the optimizer.…”

“…Then, in AlexNet-SVM, they extracted feature maps with a retrained AlexNet in order to feed an SVM classifier; they achieved the best AP of 99.64%. We tested the approach of FishResNet [ 31 ] by using the same test set, we obtained 95.62%. We used the provided code in AdvFish [ 32 ], and we trained ResNet50 by using 7-Fold cross-validation; we achieved 90.99%.…”

“…Additionally, they are higher compared with sate-of-the-art methods. We note that some methods have good accuracy, but looking at the species precision, we find that they are low for some species such as ResNet [ 30 ] for species DR, CC and AC, FishResNet [ 31 ] for species AN, ZS and NN, and DeepFish [ 7 ] for species AN and NN. We can conclude that our approach has better advantages compared with other methods.…”

“…CNNs have also shown high performance for underwater vision enhancement [ 20 , 21 ], and for fish detection [ 5 , 11 , 22 , 23 ]. For fish species classification, the first works based on CNNs applied well-known pretrained networks such as AlexNet [ 24 , 25 , 26 ], VGGNet [ 27 ], GoogleNet [ 28 , 29 ] and ResNet [ 30 , 31 , 32 ]. Our approaches use ResNeXt 101 network with 101 layers [ 33 ].…”

Underwater biotope mapping: automatic processing of underwater video data

“…For better recognition of classes, aggregated open-source data [7], and data from the White Sea are mixed in a proportion of 60 to 40. As an architectural solution, we used ResNet [8] pre-trained on 1000 classes and re-trained on the classes of interest. As a result of finetuning, this model achieved an accuracy of 95 percent on the validation set.…”