Non-Linear co-registration in UAVs' images using deep learning

Unmanned Aerial Vehicles (UAVs) has stood out for assisting, enhancing, and optimizing agricultural production. Images captured by UAVs allow a detailed view of the analyzed region since the flight occurs at low and medium altitudes (50m to 400m). In addition, there is a wide variety of sensors (RGB cameras, heat capture sensors, multi and hyperspectral cameras, among others), each with its own characteristics and capable of producing different information. In multi-spectral images acquisition, we use a distinct sensor to capture each image band and at different time, leading to misalignments. To tackle this problem we propose to train a deep neural network to predict the vector deformation fields to perform the registration between bands of a multi-spectral image. The proposed approach has an accuracy ranging from 89.90% to 93.79% in the task of estimating the displacement field between bands. With this field estimated by the network, it is possible to register between the bands without the need for manual marking of points.

Automated summary

Unmanned Aerial Vehicles (UAVs) have played a significant role in assisting and optimizing agricultural production. With the ability to capture detailed images at low and medium altitudes, UAVs provide valuable information for analysis. To address the misalignment issue in multi-spectral images, researchers propose training a deep neural network to predict deformation fields and achieve accurate registration between bands.