A novel multi-level hybrid lightweight water segmentation network called MLHI-Net has been proposed, significantly enhancing shoreline detection for unmanned surface vehicles (USVs).
Shoreline detection is pivotal for the autonomous navigation of USVs, which are increasingly being utilized for various public services like waterway inspection and quality monitoring. Most existing methods struggle with detecting shorelines due to the complex environmental conditions as well as the high similarity between shorelines and reflections, which complicates the differentiation.
The MLHI-Net addresses these challenges by leveraging advanced convolutional modules and attention mechanisms. Its architecture incorporates a convolutional module known as ORRD (Over-parameterized and Redundancy removal) along with the multi-branch dual-layer attention fusion module (MDA). The ORRD module enhances feature representation and reduces redundancies, significantly improving the network’s ability to identify water features.
Through comprehensive testing on the newly constructed CityWater dataset, which simulates various real-world conditions, MLHI-Net achieved remarkable results. The model recorded Mean Intersection over Union (MIoU) scores of 97.86% and Pixel Accuracy (PA) of 98.92%, surpassing many state-of-the-art segmentation models. This level of accuracy demonstrates its effectiveness even under adverse environmental circumstances like reflections from water surfaces and varying light conditions.
The MDA module plays a significant role by employing parallel branches for channel, spatial, and pixel attention, allowing the model to focus on different levels of detail across image features. This capability ensures both semantic accuracy and spatial fidelity, facilitating precise shoreline delineation.
The findings underline MLHI-Net's potential, offering a promising tool for improving the operational efficacy of USVs. By providing accurate water segmentation and shoreline detection, USVs can navigate more effectively, carrying out their designated tasks with greater reliability.
The continuous development of such technologies is not without its challenges. Factors like seasonal variations and the presence of small obstacles on the water surface need to be accounted for to fully optimize the navigation capacities of unmanned boats. Future work will also attempt to include segmentation of these obstacles to bolster navigation safety.
A dataset named CityWater has been constructed for validating MLHI-Net, featuring 3,215 annotated images captured under diverse conditions to allow comprehensive evaluation of the model. Notably, MLHI-Net was demonstrated to maintain its accuracy across all tested environmental scenarios, showing robustness and adaptable performance.
The MLHI-Net framework, by combining efficiency with high-performance metrics, indicates significant strides in the capabilities of USVs to operate within complex urban waterways. Given the pressing demands for environmental monitoring and public service operations, this advancement has the potential to contribute meaningfully to the deployment of autonomous systems.
Real-time operational tests conducted using this model revealed its capacity to seamlessly assist USVs by accurately delineation water edges, thereby aiding navigation and operational tasks. These promising results reinforce the MLHI-Net's contributions to the field of autonomous navigation.
Therefore, the advancements presented through this study signify not just incremental improvements but meaningful progress within unmanned vessel operations, setting the stage for possible future developments akin to refining these technologies for broader applications.