Measurement of Fish Morphological Features through Image Processing and Deep Learning Techniques

Featured Application The work described in this paper will support applications that can be employed in fish culture or in the wild. These applications can be used to monitor fish growth or health while fish classification can also be performed. The deep learning part of the application has already been developed in Python and the image processing part in Octave. Implementation of future application versions in the Microsoft Xamarin cross-development environment will allow deployment on any desktop or mobile platform. Noninvasive morphological feature monitoring is essential in fish culture, since these features are currently measured manually with a high cost. These morphological parameters can concern the size or mass of the fish, or its health as indicated, for example, by the color of the eyes or the gills. Several approaches have been proposed, based either on image processing or machine learning techniques. In this paper, both of these approaches have been combined in a unified environment with novel techniques (e.g., edge or corner detection and pattern stretching) to estimate the fish's relative length, height and the area it occupies in the image. The method can be extended to estimate the absolute dimensions if a pair of cameras is used for obscured or slanted fish. Moreover, important fish parts such as the caudal, spiny and soft dorsal, pelvic and anal fins are located. Four species popular in fish cultures have been studied: Dicentrarchus labrax (sea bass), Diplodus puntazzo, Merluccius merluccius (cod fish) and Sparus aurata (sea bream). Taking into consideration that there are no large public datasets for the specific species, the training and testing of the developed methods has been performed using 25 photographs per species. The fish length estimation error ranges between 1.9% and 13.2%, which is comparable to the referenced approaches that are trained with much larger datasets and do not offer the full functionality of the proposed method.