Feature based image registration using CNN features for satellite images having varying illumination level

Many times, meaningful information is derived from the image fusion. In applications like change detection, cancer growth detection, etc., there is a need of alignment of two or more images first. In the image registration process, two images are geometrically aligned, which is an important pre-processing step in the fields such as remote sensing, medical, etc. This paper focuses specifically on satellite images, which are commonly used in applications like change detection, weather forecasting, and growth monitoring. In these applications, image registration is a crucial step, and the accuracy of the registration process is essential. However, there are various challenges to image registration, one of them is illumination change in multi-sensor, multi-spectral satellite images. To address this challenge, paper proposes a feature based approach, where feature detection using speeded up robust feature (SURF) and descriptor from modified visual geometry group (VGG16) convolutional neural network (CNN) structure are used. The descriptors are generated from the initial convolutional layers of the modified VGG16 structure for each key point detected by SURF. The main goal of this approach is to reduce incorrect matches, which in turn improves image registration. Results of this experiment demonstrate 20% to 40% of significant improvement in correct match rate (CMR), which in turn improves image registration by the proposed approach, as compared to the method where only the original SURF is used for feature detection and descriptor generation. Therefore, it is found that, the use of CNN features as descriptor with SURF as feature detector provides improved results in terms of CMR and thus improves image registration compared to the taken method in comparison. This shows that the use of learned feature as descriptor has potential to improve the image registration.