OFRD: Optimized Feature Reweighted DenseNet for Semantic Segmentation of Complex Indoor Scenes and Marine Applications

Abstract

Semantic segmentation within an image, which involves mapping pixels to semantic labels, is crucial for various
applications, including marine applications. However, existing methods often struggle with the complexities of indoor environments, characterized by significant occlusions and clutter, which leads to suboptimal segmentation performance. Traditional deep learning models, while effective, face challenges related to computational complexity and the integration of multi-modal data, particularly when combining RGB and depth images.This paper introduces the Optimized Feature Reweighted DenseNet (OFRD), a novel deep learning framework designed to address these limitations for indoor and marine applications. The OFRD framework utilizes a DenseNet-121 backbone for robust feature extraction from RGB and depth images, effectively capturing the intricate semantics of indoor scenes. The architecture includes an RGB-D fusion stage that employs a Squeeze and Excitation module for effective feature reweighting, ensuring that important features are emphasized while irrelevant ones are suppressed. Additionally, a Region Proposal Network (RPN) is integrated to extract regions of interest (RoIs) with higher semantic relevance, facilitating precise object localization. The decoder unit processes these RoIs to produce detailed segmentation outputs, enhanced by skip connections for fine-grained results. To optimize the model’s performance, the Wild Horse Optimization (WHO) algorithm is employed to minimize the loss function, improving the accuracy of bounding box predictions. Extensive experiments conducted on the SUN RGB-D dataset demonstrate that the OFRD model significantly outperforms existing state-of-the-art methods, achieving superior pixel accuracy at 99.0%, mean precision (mP) at 95.1%, mean Dice Coefficient (mDice) at 97.6% and mean Intersection over Union (mIoU) at 95.5%.