Etudiant : LAMGARRAJ MOHAMED

Filière : Master Web Intelligence et Sciences des Données (WISD)

Encadrant : Pr. YAHYAOUYI ALI

Annèe : 2022

Résumé : Submarine environments are very rich and diversified environments, these environments attract biologists to discover their secrets, Visioon is a company of biologists who seek to discover these environments, they seek to take advantage of the power of artificial intelligence to automate stains from the observation of these environments. Their goal is to have a powerful system that can detect fish species and follow them in a video captured by underwater cameras. Deep learning and convolutional neural networks CNNs in particular, makes it possible to launch many advanced works of automatic image processing, especially to study the contribution of CNNs to the resolution of different tasks in computer vision. In the internship, we take advantage of the performance of CNN-type networks for the detection and monitoring of fish species in underwater video images. The work in internship was composed of two main phases: the detection of fish species in a first place, and the monitoring of these in an underwater video using deep learning. The contributions are summarized in the following: — First, we implement an approach for fish detection in underwater video images. This approach is based on the fusion of two parallel deep neural networks. A first network extracts the appearance features of each video frame. These features can be texture, shape and color. While the other network extracts motion features from successive images. Movement characteristics can be very relevant. The fish appears in multiple frames of a video, and can change direction and posture , which also impacts feature representation. We exploit this temporal information in addition to the appearance information to improve the detection performance. — Next, we address the problem of monitoring fish species. Deep learning requires large databases for better performance. However, the available fish image databases are small. To overcome this problem, we use the transfer learning approach in different strategies while addressing various issues (choice of color space, elimination or not of background, and way of artificial data augmentation). ).