The state of implementations is currently in a typical research state, far from usability in a real product. But it oers a good perspective on future extensions and research. Many improvements are already highlighted in the chapters of this thesis. In the following, we are summarising some of the key areas with an outlook of potentially more high level improvements.
An obvious one is, to make the implementation robust to changes beyond the current limitations, mainly for example in cases of to strong, sudden light changes or changes in the camera's perspective. For scenarios, in which we cannot build on a priori knowledge or distinct reoccurring objects that can be tracked, the method for the scene segmentation becomes very important. We can expect strongly improved accuracy of tracked colour data, if the scene is segmented with sucient precision. The segmentation should result in a separation of distinct objects or contiguous surfaces of a mostly homogeneous colour appearance.
The implemented proler in this work still oers vast potential for improvement. The points for this are distinctly in the quality of the spline and curve tting, as well as in a generation of pre- and post-linearisation curves. Also an optimiser can be useful, that balances the tting of the per-channel curves with the spline tting of the CLUT.
We are currently only using an ane transformation as a corrective colour transforma- tion. Other models can potentially be more suitable in general, or for particular cases. Dierent models and means of parameterising them can be derived and evaluated. Ideally, such a model could evolve, rather than being rigidly specied. For example neural nets or genetic algorithms could be used to determine a distortion model or adaptation parameters for a transformation.
A hybridisation of for example the White Patch Retinex algorithm with time based colour sample tracking seems like a promising modication. Bright specularities and black
11.4. FUTURE RESEARCH POTENTIAL 159 points can often be found easily within scenes. In the presence of these, they can be used in two ways. On the one hand, to provide further data points. But on the other hand, these distinguished pieces carry more information as we know that they dene the neutral axis, and as such, they are very useful to derive better colour distortion model parameters.
Finally, as mentioned in this thesis, the processes of deriving a colour correction transfor- mation and of colour correcting are only very loosely coupled. Therefore, the implementation can easily be decoupled and potentially distributed, for example by using a service oriented architecture (SOA).
Chapter 12
Conclusions
We have built a system that perceives colour and does that as precisely as possible while adapting to light changes, when encountering common natural and articial light sources. It obtains device independent and visually linear colour descriptors (CIE LAB colour space). The motivation for using this colour space is driven by its benecial properties for analysing and comparing colours. Therefore, we have also based the time adaptive colour correction process to operate on this colour space.
For these reasons, the rst step to this work was a thorough introduction to the pro- cess of colour capturing and perception, along with the encoding of dierent colour spaces. The dierent involved colour spaces were described in their origin, and compared by their individual properties, to foot the undertaken research on a solid base in colour science.
As a verication that colour corrective operations are sensible on CIE LAB colours, we have modied the well known White Patch Retinex and the Grey World Assumption colour constancy algorithms. Results obtained from these alterations were successfully validated using Colour Indexing. In this process, the eectiveness of the colour constancy algorithms operating on device RGB and CIE LAB colour spaces were subject to a comparative analysis using identication rates for various objects.
The approach for adaptive colour correction taken considers the additional factor time, which could be used to extract further input data. This is in opposition to colour constancy, which only bases corrections on one image at a time. Information gathered by including time as a dimension was two fold: Primarily, a sequence of images in time enables us to study the shift of coloured elements in the scene; But also the fact that certain scene elements only change slowly can be taken into consideration, to ease the burden of having to adjust a colour correction for every frame individually. We can sum up, that the rst enabled us to adapt colour corrections for the given moment; and the second made it possible to only update the correction in intervals. Thus, real time processing of corrected video feeds becomes possible.
The corrective process to generate an updated ICC prole is outlined. Due to the above mentioned common slowness of colour changes due to illumination, we can run this prole adaptation decoupled from a capturing process. The capturing process itself can operate mostly independently, and just receive occasional updates for the current ICC prole, to apply it to the incoming image stream from the camera (using a colour management system), enabling it to operate in (near) real time conditions.
12.1 Contributions
We have been able to show colour constancy can with minor modications also be based on derived opponent colour spaces, as opposed to raw device colour spaces. In our research, we have employed processing on the device independent, linear CIE LAB colour space for this. This choice was fueled by the attempt to fuse practices of credible colour processing with various common approaches of colour correction. Besides the already mentioned colour constancy, these were inspired by dierent basic directions of correction attempts.
Firstly, the application of ICC proles through a CMS can yield a good quality colour correction for xed conditions, if the prole corresponds to the given conditions. The ap- plication of ICC proles forms a proven, standardised, much applied base for additional corrections.
Further ideas were taken from biological and probabilistic approaches. Objects of interest are viewed in the context of a scene, which can be used to derive additional information for a correction. This information can be correlated with corresponding information acquired from previous or reference frames, to gain an understanding of the encountered shift in colour appearance through a temporal element beyond a single frame. A best t matching of parameters for a colour space distortion model (through an approximation by an ane transformation) is performed, to derive an approximate correction for increased colourimetric stability.
Finally, known references within the scene's objects or context were harvested for the purpose of gathering colourimetric information, that can be used to derive the correction. This was achieved by tracking (reoccurring) objects or other scene elements. Their colouring then could be matched to the appearance from previous or reference frames. A priori knowledge can play a major role in this, where an application can build on distinct knowledge of appearance or composition of objects and scenes.
A key step in achieving this goal, was to move from single frame corrections (as used in colour constancy) to a time enhanced colour correction method (deriving information from a consecutive series of frames). These multiple corrective contributions were then fused into one standardised ICC prole, for a simple corrective ICC transformation, which is valid for the current illumination case only. This nal, comprehensive input device prole closes the loop towards easy and well dened corrective colour acquisition with capturing devices.