A Method for Constructing an Aerial Image Dataset of Civil Aircraft “Black Box” Using AIGC-Generated Samples

Using drones for automatic cruise to capture crash sites and employing AI algorithms to automatically detect aircraft wreckage and the “black box” can significantly improve the efficiency of onsite search and rescue. To address the issue of insufficient “black box” image samples from crash sites, this paper proposes a method for constructing an aerial image dataset of “black box” using Artificial Intelligence Generated Content (AIGC) for background generation and target sample synthesis. First, multi-angle images of the “black box” and background images in typical terrains are collected using drones, and the actual aerial image samples are formed after segmentation and labeling. Next, a virtual aerial background image generation method based on AIGC is proposed. This method uses ChatGPT interactive prompts and large image generation models to generate aerial background images in complex terrain environments. Finally, using collaborative dual-transform automatic image synthesis technology, segmented real target images are randomly embedded into background images, and virtual aerial samples are coordinated and generated, thereby constructing the “black box” target detection dataset. This dataset achieved good training and testing results on five classic target detection models, and experiments verified its effectiveness. In the ablation experiments on various subsets using YOLOv8n, the detection accuracy on the real sample set improved from 0.76 to 0.976 on the combined sample set, with an accuracy increase of 28.4%.