RESULTS: Image Blending
Image blending is the final step in producing high quality image mosaics. Radiometric variations in overlapping views and violation of certain scene assumptions such as rigidity, stationary, and (or) planarity, lead to geometric misalignments (e.g. due to parallax effect) and photometric differences. Upon blending, these usually result in degrading artifacts, such as blurry regions or artificial seams.

Sources of Misalignment
The concept of misalignment concerns the presence of features (i.e. points, edges or object shapes) not matching perfectly within two stitched views of the same scene. Misalignments in registered images are common, even when a perfect match (in terms of matched features and outliers rejection) has been obtained.

There are four main reasons that make misalignments noticeable in the seams of the registered images. From the geometrical point of view, non precise registration, planarity assumption viola- tion and moving objects are the most relevant. From the photometrical point of view, changes in the location of the illumination sources and in the exposure time are the principal reasons.

First cause of misalignment is a non precise registration. These imprecision may be due to a non enough accurate finding of features and correspondences into the image pair. Another reason may be a numerical inaccuracy when solving the equation system that allows estimating the homography matrix. In any case, if registration is not good enough, the visibility of the seams between the implied images will be evident. Parallax effect appears when the planar assumption constraint is violated, becoming a difficult problem to solve in the overlapping and seam regions. The regions of the images becoming to an object presenting parallax may be ignored in the motion estimation procedure only if they are not enough significant relative to the size of the frames. Nevertheless, the misalignment will still being existent. It will be necessary to determine which image will contribute to the mosaic with the information relative to the conflictive object.

Illumination variations and non-uniform illumination can be a cause of misalignment materialized in noticeable inconsistences in the seam. This kind of issues mainly appear when artificial lighting is used in order to illuminate the scene.




Underwater mosaic showing evident geometrical and photometrical misalingments rendered using a single "last-in" stitching, before the application of any image blending techinque.




Underwater mosaic after the application of a hybrid image blending technique.


Blending Techniques
Nowadays there are two main approaches to image blending in the literature: transition smoothing and optimal seam finding. On the one hand, transition smoothing methods (also known as feathering or alpha blending methods) attempt to minimize the visibility of seams by smoothing the common overlapping regions of stitched images. On the other hand, optimal seam finding methods place the seam between two images where intensity differences in their common overlap- ping area are minimal. Furthermore, it is also possible to find techniques that take profit of the benefits of both approaches, in what can be called hybrid techniques.

A new proposed hybrid method wants to solve both registration and illumination problems, targeting the efforts in reducing them in underwater image sequences, taking into account their particular properties. Due to limitations in the underwater lighting systems, illumination variations between frames and non-uniform illu- mination are frequent in image sequences where camera is pointing to the seabed. The planarity assumption of the scene depends on the distance between the camera and the seafloor, being pos- sible to consider this one as planar if the mentioned distance is significantly higher than the relief of the seafloor elements. Nevertheless, in an unexplored underwater terrain the assumed condition may be easily violated. Moving objets are materialized in fishes or plants (depending on the kind of life present at the exploration depth).




Underwater color mosaic showing evident geometrical and photometrical misalingments rendered using a single "last-in" stitching, before the application of any image blending techinque.




Underwater color mosaic after the application of a hybrid image blending technique.


Final Thoughs
Underwater imagery presents a set of characteristics determining to a large extent the kind of blending methods that may be used in order to build the mosaics. Since planarity conditions of the explored seafloor areas may not be ensured, techniques able to deal with parallax and moving objects must be considered. Furthermore illumination conditions are usually poor and dealing with photometric misalignments becomes a constant. The goal of building visually pleasant image mosaics useful for visual inspection has ben achieved.