Rybí biometrie s využitím metod počítačového vidění

Nowadays, individual fish identification is made by tagging. Tagging is an invasive method of fish identification that can cause injuries to fish and stress to them, leading to increased mortality. Tagging is a time consuming and expensive identification way. The disadvantages of this method are obvious. To solve those problems, identification from the images could be used. Image-based fish individual identification is an excellent alternative. It is cheap, fast and not stressful to the fish. Fish identification from images is widely used in species identification. But not many studies deal with the identification of individual fish. All described reasons motivate us to work toward individual fish identification from the images as a substitute for fish tagging. We have done complex research with different fish species, different data collection conditions, and long-term perspective identification. In chapter 2, we tried the first attempt to automatically identify individuals of ornamental fish Sumatra barb Puntigrus tetrazona in an aquarium. Fish were freely moved in an aquarium with water; the green background was used to do the fish segmentation. Totally 43 individuals were used in this experiment. Identification accuracy was 100% and supported us to continue with the next step experiment. The next step was to increase the number of photographed fish to 330 fish. We used the commercially important Atlantic salmon Salmo salar in this study. We have tested different visible patterns on the fish body, such as dots on the body (chapter 3) and the iris of the eye (chapter 5). The duration of the experiment was six months. In this study, different data collection conditions were tested. Images of fish underwater in an aquarium and out of the water in a photo tent were taken. For the pictures of the fisheye, we used a micro-camera. The best results obtained from those experiments were 100% accuracy for the dot approach and HOG parametrization methods and 95% for fisheye data. The last experiment in this dissertation (chapter 3) was done to prove that the fish species which have no obvious pattern on the body, such as Sumatra barb (black vertical stripes) and Atlantic salmon (dots on the body), could be identified non-invasively from the images. European seabass Dicentrarchus labrax and common carp Cyprinus carpio were used in this study to prove this idea. Totally 300 seabass and 32 carp were photographed out of the water to get high-quality data. Together with the short-term experiment, we collected long-term data (two months for seabass and four months for carp). Different parts of the body were tested for identification (lateral line, scale pattern, operculum). Surprisingly, the identification results were high enough (100% for both species, even for the long term experiments) to conclude that the photo-identification works with species without an obvious pattern on a body. The conclusion supported by the results of the experiments is that the automatic photo-identification of individual fish is possible using machine vision. Data processing and identification procedure were fully automated. The approach works for species with and without an obvious pattern in the body (Sumatra barb, Atlantic salmon, European seabass and common carp), and it is useful for long term individual identification. The method can be used as a substitute for invasive fish tagging