A Cognitive Sample Consensus Method for the Stitching of Drone-Based Aerial Images Supported by a Generative...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Drone-based aerial imaging for panoramic generation suffers from image quality degradation due to unstable shooting angles.
• Existing methods like RANSAC can produce falsely estimated hypotheses, impacting the stitching process.

Purpose of the Study:

• To propose a deep learning-based outlier rejection scheme to enhance the quality of drone-based panoramic images.
• To reduce falsely estimated hypotheses in image transform estimation for improved stitching.

Main Methods:

• Utilized generative adversarial networks (GANs) for outlier rejection in image stitching.
• Employed RANSAC with scale-invariant feature transform (SIFT) descriptors to generate training data.
• GAN's discriminator pre-judges RANSAC's hypothesis, with the generator confirming authenticity.

Main Results:

• The proposed GAN-based method effectively rejects outliers caused by unstable shooting angles.
• Demonstrated stable and good performance on drone-based aerial and miscellaneous image datasets.
• Achieved improved accuracy in image transform estimation compared to baseline methods.

Conclusions:

• Deep learning-based outlier rejection using GANs significantly improves panoramic image generation from drone footage.
• The method offers robust performance, particularly in challenging imaging scenarios.
• This approach provides a reliable solution for enhancing aerial image stitching quality.