Line of action:

Status:

Execution date:

The general objective of the project is the design, development and implementation of an artificial vision system based on deep learning capable of identifying and measuring specimens in color images obtained in fish markets.

The specific objectives of the project are in accordance with the objectives of the second cycle of the Marine Strategy, specifically with respect to objective A.L.3. Maintain or recover the natural balance of the populations of key species for the ecosystem, and also in relation to objective C.L.9., to promote that fish stocks are properly managed, so that they are kept within safe biological limits, paying special attention to those whose status is unknown, and those that do not reach the BEA according to the initial assessment of D3 in the Levantine-Balearic marine demarcation. The specific objectives defined for the project are:
1. Establish the requirements of the system
2. Preparation of a dataset of target species for artisanal
fishing tagged with metadata on species typology and size.
3. Development of a deep network architecture for specimen identification and
measurement.
4. Development of the vision system for the classification and measurement of
species.
5. Implementation of the vision system in fish markets of the Natura 2000
Network 6. Disseminate project results

A.1. Analysis of requirements and functional specification.
A.2. System design and laboratory experimentation.
A.3. Implementation and validation of the system.
A.4. Dissemination.

The Deepfish project has achieved the objective of designing, developing and implementing a deep learning-based machine vision system capable of identifying and measuring specimens in colour images obtained in fish markets. The results of this project are the following:

Dataset: A dataset of fish tray images has been produced with specimens labeled at the pixel level, along with information on the species and different size measurements of each fish. A total of 1,227 tray images were collected, with 8,245 specimens of 60 different species.

YOLACT-based deep network architecture: A deep network architecture has been developed, achieving success rates in the identification of the species close to and/or greater than 77% for 9 of the 15 species with the most individuals. The estimation of the size based on the correct identification of the species reaches errors of less than 5%.

Implanted system: The vision system for the monitoring of species and sizes has been implemented in the “El Campello” fish market of the Cabo de Huertas SCI. The tests of the operation of the recognition and carving system carried out in a real environment in the fish market have made it possible to verify success rates similar to the laboratory system.

Diffusion: Dissemination activities have been carried out both in the scientific and general fields. In the scientific field, it has participated in two international scientific congresses and two scientific publications of impact have been produced. With regard to general dissemination, two pieces of information have been published in newspapers, interviews in national media, dissemination on the web and social networks
, as well as a day of presentation of results.

The results can be found on the project website, specifically in the results section:

• Informative video of the project
• Explanatory video of the project’s dataset
• Dataset available with open access
• Explanatory video of the project architecture
• Demonstration video of the system in the fish market
• Press releases and media presence