Augmenting Naval Ship Images for Viewing Distance using Adobe Generative Fill 

Status: Ongoing |  Funding: Office of Naval Research

Image augmentation plays a crucial role in enhancing the performance of AI models, particularly in domains with limited or highly specialized datasets. This abstract explores advanced image augmentation techniques to address the challenges of zooming out images while preserving image integrity. This task is complex because coherent backgrounds must be synthesized while maintaining the structural integrity of the primary subject. Techniques were applied to a diverse pool of 98 images of US Navy ships, aiming to assess the robustness of AI tools capable of accurate image recognition across various zoom levels.

The methodology began by exploring several state-of-the-art generative tools, including custom GPTs from OpenAI, MidJourney, Google Vertex AI, Meta LLaMA, and the Generative Fill feature from Adobe Photoshop. Comparative analysis revealed that Photoshop’s Generative Fill provided the most realistic and contextually consistent results. Before applying generative fill, images were pre-processed using OpenCV to create zoom-out variations with factors of 0.5, 0.25, and 0.125. The images were centered on a white background, producing three distinct datasets for each zoom-out factor. This pre-processing step enhanced the generative model’s ability to synthesize naturalistic surroundings.

Two approaches were implemented for the Generative Fill process: one using the prompt "sea" to guide background generation and another without any prompt for a more neutral fill. Additionally, two operational modes were tested: batch processing for efficiency and manual iteration for precision. These approaches collectively produced three distinct datasets for each zoom-out factor, enabling a comprehensive analysis of augmentation effectiveness.

To evaluate image quality, a color-coded labeling system was employed—red for poor, yellow for medium, and green for high-quality images. The evaluation criteria focused on five key aspects: Subject Integrity (preservation of the ship's structure), Subject Dimensions (maintenance of accurate proportions), Background Quality (visual coherence with the original), Background Consistency (contextual harmony with the subject), and Background Artifacts (absence of unnatural modifications). The results will be validated using inter-rater reliability measures, ensuring a thorough and unbiased assessment.

This work contributes to the broader field of AI in Practice: Impacts, Risks, and Opportunities by demonstrating the potential of advanced image augmentation to enhance model performance in real-world applications. Improved zoom-out image generation enhances AI's ability to recognize objects in complex, variable-scale scenarios, which is critical for defense, surveillance, and maritime navigation systems. However, this also raises considerations regarding the ethical deployment of such AI models, particularly in sensitive military contexts. This research illustrates the balance between leveraging AI’s capabilities and addressing potential risks associated with data manipulation and bias. By advancing image augmentation techniques, this research not only enhances model accuracy but also contributes to the ongoing dialogue on responsible AI innovation. 

Contributors

Resources

Progress Timeline

• 2024-Present