Alignment of Multi-Camera Spectral Images Using Wavelet Transform

Multispectral imaging is a valuable tool for monitoring plant health in agriculture and forestry. These methods are based on the calculation of different vegetation indices consisting of at least two spectral channels. In order to obtain accurate results, these two images must perfectly overlap. For cost-effective multispectral monitoring of plants with several single-channel cameras, we propose an image alignment technique that utilizes the structure of the scene itself and edge detection by means of wavelet transform. By transforming the spectral images into the frequency domain, we can calculate the average displacements of the scene by determining the distance between the peak amplitudes of the two spectral channels for each image line and axis. Originally, we used the haar wavelet function to extract the frequency domain, which yielded initial good results for vegetation and structures. Other wavelet functions were tested, which allowed further improvements. However, compared to edge extraction with the non-wavelet Sobel operator, only the rbio3.1 wavelet showed improved capabilities in merging the spectral channels, requiring on average one iteration of the proposed algorithm, while maintaining the success rate of Sobel.