Transferable Deep Learning Model for the Identification of Fish Species for Various Fishing Grounds

Tatsuhito Hasegawa,Kei Kondo,Hiroshi Senou

DOI: https://doi.org/10.3390/jmse12030415

IF: 2.744

2024-02-27

Journal of Marine Science and Engineering

Abstract:The digitization of catch information for the promotion of sustainable fisheries is gaining momentum globally. However, the manual measurement of fundamental catch information, such as species identification, length measurement, and fish count, is highly inconvenient, thus intensifying the call for its automation. Recently, image recognition systems based on convolutional neural networks (CNNs) have been extensively studied across diverse fields. Nevertheless, the deployment of CNNs for identifying fish species is difficult owing to the intricate nature of managing a plethora of fish species, which fluctuate based on season and locale, in addition to the scarcity of public datasets encompassing large catches. To overcome this issue, we designed a transferable pre-trained CNN model specifically for identifying fish species, which can be easily reused in various fishing grounds. Utilizing an extensive fish species photographic database from a Japanese museum, we developed a transferable fish identification (TFI) model employing strategies such as multiple pre-training, learning rate scheduling, multi-task learning, and metric learning. We further introduced two application methods, namely transfer learning and output layer masking, for the TFI model, validating its efficacy through rigorous experiments.

oceanography,engineering, marine, ocean

What problem does this paper attempt to address?

The problem that this paper attempts to solve is: **Automated fish species identification and length measurement to promote the efficiency and accuracy of fishery resources surveys**.

Specifically, in current fishery resources management, the manual recording process of detailed fishing information (such as species, length, quantity, etc.) is very cumbersome and time - consuming. With the increasing demand for sustainable fishery development, it has become particularly important to automate the processing of fishing information. However, existing image recognition systems face the following challenges when applied to fish species identification:

1. **Fish species diversity**: There are a large variety of fish species caught in different seasons and locations, and it is difficult to cover all possible fish species with a single model.
2. **Data scarcity**: There are relatively few publicly available large - scale fish image data sets, making it difficult to train a general - purpose model.
3. **Difficulty in fine - grained identification**: Some fish are very similar in appearance, and distinguishing them is a complex task.

To solve these problems, the author proposes a **Transferable Deep Learning Model (TFI)** for identifying fish species in various fishing grounds. This model is pre - trained on a large - scale fish image database and can easily adapt to different fishing ground environments through transfer learning or output - layer masking techniques. The following are the main contributions of this method:

- A large - scale fish identification data set (KPM) containing approximately 200,000 images was constructed, covering 2,826 fish species.
- Experimental results show that when tested using public data sets, the TFI model has better transfer performance than a model pre - trained only with ImageNet.
- In a subset of the Japanese fishery resources survey, without additional training, the model achieved an identification accuracy rate of 72.9% for 129 fish species.

Through this method, researchers hope to achieve an efficient and accurate fish species identification system, thereby improving the efficiency and scientific nature of fishery resources management.