Doctorate thesis defense of Walid Messaoud

Attention Mechanism for Deep Learning in the Analysis  of Maritime Environment

Presented by

Walid Messaoud

On 14/12, 2024 at .09.H, in Amphitheater Ibn Roched, SUP'COM 2.

Thesis Committee

Abstract

This research introduces the pioneering progress of object detection in the maritime field, to address its distinct challenges and push technological boundaries to enhance security, ensure navigation safety, and preserve the environment. The complexity of the maritime setting, characterized by a variety of dynamic objects, diverse vessel outlines, occlusions, and constraints within annotated datasets, presents distinct challenges. In response, we undertake a groundbreaking implementation of attention mechanisms as a key solution, directing attention to crucial image regions and adapting to the ever-changing conditions of the maritime setting.

We lead the way in applying Multi-Head Self Attention (MHSA) in the maritime realm, conducting a comprehensive evaluation of various maritime datasets to identify the most suitable ones. This pioneering effort represents a significant leap forward in improving object detection within the maritime environment. Our evaluation covers six cutting-edge models that incorporate attention mechanisms and the Low-Visibility Enhancement Network (LVENet). Utilizing maritime datasets such as SeaShips and the Singapore Maritime Dataset (SMD), our approach demonstrates substantial improvements in mean Average Precision (mAP) scores.

LVENet, designed specifically to address challenges in scenarios with reduced visibility, significantly reinforces object detection. This improvement leads to a varying mean precision improvement across a spectrum of models. Our comprehensive assessment not only marks a pivotal achievement in maritime object detection but also underscores advancements in both precision and simplicity, advocating for enhanced deployment strategies for these models in real-world maritime applications.

Furthermore, we highlight the effectiveness of a proposed adaptive pruning approach, derived from Layer-wise Relevance Propagation (LRP) in attention mechanism-based models: Bottleneck Transformers (BoTNet) and Attention Augmented Convolutional Networks (AACN). Our experiments encompass benchmarks, such as ImageNet, Pascal VOC, Canadian Institute for Advanced Research-10 (CIFAR-10), and Common Objects in Context (COCO) datasets, confirming the versatility of our approach. Notably, we extend our assessment to the maritime domain, proving the effectiveness of adaptive pruning in models tailored for maritime applications, specifically BoTNet and AACN. This validation showcases an average epoch time reduction of over 30\%, preserving accuracy and highlighting the adaptability of our proposed method in enhancing computational efficiency for maritime-related tasks.

In addressing the limited datasets in the maritime domain, our exploration delves into the realm of GANs for latent space manipulation. Our proposed method leverages a sophisticated framework that integrates a classifier, auxiliary mapping composed of transformer blocks, and the implementation of conjugate gradient. This novel approach enables precise attribute modification, showcasing its efficacy in manipulating latent spaces. Subsequently, our research presents compelling results in facial attribute manipulation using FFHQ and CelebAHQ datasets, demonstrating remarkable superiority by exceeding the state-of-the-art approaches in terms of Manipulation Disentanglement Score (MDS). Extending our method to the maritime domain, we pioneer advancements by surpassing the state-of-the-art Surrogate Gradient Field (SGF) method. Specifically, our GAN-based approach achieves substantial improvements using SeaShips and SMD datasets. These results highlight our method's remarkable strides beyond the existing benchmarks, illustrating its unparalleled efficiency in both facial attribute editing and maritime image manipulation. Additionally, we prove the versatility of our approach by conducting extensive experiments with different versions of StyleGAN, showcasing its adaptability across various GAN architectures. Our proposed approach pioneers GAN's generation of real-world images in the maritime environment, establishing new frontiers in the application of GANs.