Convolutional Versus Operational Neural Networks for Semantic Segmentation of Maritime and Harbor Scenes
10 Pages Posted: 12 Jun 2023
Abstract
Semantic segmentation aims to associate each pixel of an image with a corresponding label that describes what is being depicted. The key aspect of semantic segmentation emanates from incorporating the classification of objects and the recognition of their shape, which is paramount for the autonomous transportation industry, e.g., maritime autonomous ships and self-driving vehicles. Various semantic segmentation solutions are based on standard Convolutional Neural Networks (CNNs). Nevertheless, recent evidence suggests that self-organized operational neural networks (Self-ONNs) can yield better performance because of their increased heterogeneity and learning capacity. This paper presents a novel network approach by combining convolutional layers with operational layers for segmenting objects in a maritime/urban environment. Overall, the findings show that the operational layers are compatible with convolutional layers for semantic segmentation tasks. While the ResNet-18 model with convolutional layers achieved 19.8%, our model SelfONN-18_2 achieved 25.4% in Mean intersection-over-union (Mean IoU) over the former validation set of ADE20K Dataset. Even with a single layer of operational layer, we achieved better results. While PSPNet-18 with convolutional layers achieved 21.3%, our model SelfONNet-18 with a single operational layer achieved 32.5% in Mean IoU over the former validation set of ADE20K Dataset.
Keywords: semantic segmentation, Operational neural networks, Self-ONNs, Generative neurons, Machine Learning, Maritime scenes
Suggested Citation: Suggested Citation