Towards devising an effective and reliable means of fish detection and classification through the exploration of various deep learning algorithms

Abstract

Due to a number of reasons, marine ecosystems change with certain species of fish disappearing while novel species of fishes become a new staple within a given ecosystem, e.g., a lake, river, etc. Monitoring these changes in ecosystems as different species dwindle and swell in number is crucial for marine researchers, fishery owners, and fish species preservation programs. These increase and decrease in numbers indicate changes in environmental conditions that either favours a certain species or does not. In order to study these changes in conditions, it is imperative to firstly detect the changes in the population of species which is where we come in. The challenges for an underwater project range from water pressure, lack of sunlight, different orientations of fish, the motion of aquatic plants, riverbed structures, and the sheer diversity of shapes in different species. Machine learning and image processing technologies can be of significant importance in identifying such underwater fish species. In our research, we decided to use Convolutional Neural Networks (CNN), namely YOLOv4, to detect fish in input image frames. To classify the fish species, we will use a CNN network. The fusion of these networks is proposed in order to achieve a high level of classification accuracy of fish species from smallsized samples. In order to demonstrate the effectiveness of the model, we propose two datasets, namely BDIndigeneousFish and A-Large-Scale-Fish-Dataset is used, which contain a vast range of image data of several species from different habitats. The image data is fed into the Darknet, which identifies and detects the fish pixels in the image frame. Furthermore, these input images are then passed on to CNN for classification.