Enhancing Maritime CV with Domain Knowledge

By Dr Silpaja Chandrasekar, PhDReviewed by Susha Cheriyedath, M.Sc.Aug 27 2024

In a paper published in the journal Applied Sciences, researchers introduced innovative methods to enhance computer vision (CV) techniques in the maritime industry. They focused on improving detection and recognition accuracy through ensemble learning and domain knowledge integration.

They also proposed a novel application of CV combined with transfer learning to optimize online video viewing on ocean-going vessels, addressing challenges related to low-speed, high-cost internet services. This research broadened the scope of CV applications beyond traditional water surface target recognition, offering significant advancements for the shipping industry.

Background

Past work in the maritime industry has extensively applied CV techniques for naval surveillance, safety, and environmental protection, leveraging advancements in artificial intelligence (AI), big data, and the Internet of Things (IoT).

While deep learning (DL) methods like region-based convolutional neural networks (R-CNN) and you only look once (YOLO) have improved detection accuracy, challenges persist due to complex water surface environments that degrade image quality, impacting detection reliability. Despite advancements in visibility enhancement networks, achieving flawless image restoration and consistently accurate detection in maritime CV models remains challenging.

Enhancing Maritime CV Accuracy

In the shipping industry and maritime research, improving the accuracy of CV models for detecting and recognizing water surface targets is crucial. Traditional methods often need help with environmental challenges, such as fog and light reflection, leading to inaccuracies in ship recognition.

For example, cameras installed at key maritime locations might mistakenly identify a sea vessel as a river vessel, causing inefficient management. Two innovative methods are proposed to address this: enhancing recognition accuracy through ensemble learning and integrating shipping domain knowledge into the recognition process.

Ensemble learning combines predictions from multiple models to improve accuracy beyond a single model. Installing cameras at different angles allows CV models to process ship images from various perspectives, with the final output based on the correct predictions. For example, nine cameras with 70% accuracy each can achieve an overall system accuracy of about 90.12%, enhancing recognition reliability.

Integrating shipping domain knowledge into CV models can significantly improve recognition accuracy by accounting for practical maritime conditions. Understanding tugboat behavior helps refine models to detect inconsistencies, such as identifying a tugboat moving unusually fast or slow.

These innovative methods—ensemble learning and domain knowledge integration—represent significant advancements in CV model accuracy for the maritime industry. By addressing the limitations of traditional approaches and leveraging the strengths of multiple learners and practical maritime insights, these techniques offer more reliable and effective solutions for detecting and recognizing water surface targets, ultimately improving the safety and efficiency of marine operations.

Optimizing Maritime Video

A novel application of CV techniques in the maritime domain aims to enhance ocean-going vessels' online video-viewing experience while conserving network resources. Due to the reliance on communication satellites, network speed on these vessels could be faster, and charges are high.

Crew members, often at sea for weeks or months, require efficient Internet access for online videos. Existing solutions involve compressing and decompressing video files to reduce transmission load, but this process is time-consuming and can degrade video quality.

The proposed solution leverages a CV model and transfer learning to improve video quality while minimizing data usage. Transfer learning allows knowledge from a pre-trained model to be adapted to new, similar tasks.

Specifically, a CV model is pre-trained to transform low-resolution videos into high-resolution formats. When crew members watch videos, the software loads the initial 10% of the video in high resolution and the rest in low resolution.

The high-resolution portion is used to fine-tune the pre-trained CV model, adapting it to the specific video type. The optimized model then converts the low-resolution remainder into high-resolution, creating a final video that combines both portions, enhancing the viewing experience while saving data.

This approach offers a significant reduction in data usage. For instance, a high-resolution movie that requires 2 GB of data only requires 0.5 GB using this method, with a similar viewing experience. The entire process, including video loading and model fine-tuning, is automated within the software, providing a practical and efficient solution for online video consumption on ocean-going vessels.

Conclusion

To sum up, CV techniques were widely applied in the shipping industry and maritime research, though challenges like light reflection and adverse weather often compromised image quality and target recognition reliability. This study introduced innovative methods for enhancing CV model accuracy, including ensemble learning and integrating domain knowledge.

A novel application of CV techniques also addressed low-speed, expensive network services on ocean-going vessels by improving online video viewing while conserving resources. Future research should further enhance CV model accuracy and develop practical solutions for video viewing challenges on vessels.