Summer school

Project name

Fish detector training by artificial 3D fish model

Leader

Ing. Petr Císař, Ph.D.

Description of the activities of the project

• Fish behavior is one of the indicators of welfare and reflects the overall condition of fish, including the effects of diseases and rearing conditions. In recent years, computer vision has been increasingly used for behavior analysis, performing detection of individual fish and analyzing their movement in rearing tanks. This approach is one of the methods of so-called precision aquaculture, which aims to automate fish production.
• Fish detection in video recordings is carried out using image processing methods, specifically convolutional neural networks. The detection quality of these methods directly depends on the quality and quantity of the data used to train the neural network. These training data must contain a large number of objects (fish) in various orientations, sizes, deformations, lighting conditions, backgrounds, and water transparencies. Acquiring such data is time-consuming and requires extensive data recording. The subsequent training process is then based on manual annotation of fish positions in images. The neural network uses these data and annotations for training. This process is both time- and cost-intensive.
• For training purposes, so-called synthetic data can also be used; these data are generated entirely by a computer and represent a simulation of the real world based on knowledge of how it functions.
• The aim of this project is to create a 3D model of rainbow trout that, in terms of appearance and deformability, corresponds to the real appearance and movement of the fish. To create the model, it will be necessary to analyze the range of deformability of individual parts of the fish (head region, tail region, fins, etc.) during real swimming in a tank and to create digital models of several individuals. The model will then be used to generate synthetic images of fish in a tank for training neural networks. The outcome of the work will therefore be a digital fish model and a description of its deformability. Knowledge of image processing or 3D modeling is not required to solve this task. The model will be used to generate the trainind data for training of neural network and the network will be tested on the real video record of the fish in tank.
• The students will learn how to create 3D model, how to train and use neural networks for fish detection and how to perform experiments with the fish.

Students

• Number of students - 1-2
• high school students students - NO
• university students - YES
• Prepositions (what should the students know): basics of programing