PBIM: A Patch-Based Image Mosaicking for UAV Images with Natural Scene Distortions

The process of creating a seamless mosaic from a set of UAV-acquired images involves aligning and blending individual images to produce a coherent and comprehensive representation of the surveyed area. However, images captured by drones often suffer from distortion, specifically caused by the complex interaction between the camera sensor, lens, and the three-dimensional scene-referred to as natural scene distortion. These distortions significantly impact the overall visual quality of the generated UAV mosaic. While this problem has been extensively studied in the context of conventional photography, it has received less attention in the context of UAV -based images. The purpose of this study is to develop a less complex solution comprising two components: 1) patch-based image mosaicking (PBIM) and 2) CNN-based distortion detection and classification via transfer learning. The patch-based image mosaicking outperformed other existing software solutions in terms of creating a comprehensive visualization of the UAV mosaic. On the other hand, the CNN-based component identified the best model with a pre-trained VGGNet network, achieving a highest accuracy of 65.59% in detection and classification, as well as an average of 10.25 in the generation of sub-directories for data aggregation across the entire test dataset. The PIQE evaluation metric indicated that the majority of the U A V mosaics produced using the proposed solution exhibited good quality. Furthermore, when assessed using the background proportion metric introduced in this study, the proposed method produced the best overall visual appearance quality compared to other solutions.