This chapter addresses the problem of processing hyperspectral images (HI) and sequences leading to high efficiency implementations. A new methodology based on the application of cellular automata (CA) is presented to solve two different processing tasks, the segmentation and denoising of HI and sequences, respectively. CA structures present potential benefits over traditional approaches since they are computationally efficient and can adapt to the particularities of the task to be solved. However, it is necessary to generate an appropriate rule set for each particular problem, which is usually a difficult task. The generation of the rule sets is handled here following a new methodology based on the application of evolutionary algorithms and using synthetic low-dimensionality images and sequences as training datasets, which results in CA structures that can be used to process HI and sequences successfully, thus avoiding the problem of lack of labeled reference images. Both processing approaches have been tested over real HI providing very competitive results.
TopIntroduction
In hyperspectral images, the spectral information of every pixel is collected in a large number of contiguous discrete spectral bands. The wealth of information provided by the large amount of data produced for a single scene is a great help in solving a variety of processing tasks. However, practical hyperspectral applications typically require these large amounts of data to be processed in (near) real-time. Decreasing the data processing time involves, on the one hand, the design of time-efficient data analysis techniques. On the other hand, it is desirable that these algorithms can be easily processed in a concurrent fashion within hardware such as GPUs
In this respect, Cellular Automata (CA) are some of the most common and simple models of parallel computation. CAs are dynamic systems consisting of a regular spatially distributed grid of cells, each one characterized by its state. The state of every cell is updated in parallel depending on its current state, the state of neighboring cells and a set of transition rules. The crucial point in the use of CAs is to properly determine this set of transition rules, that is, to infer a set of rules that when applied locally to every cell, lead to the desired global behavior of the automata, which is far from being a straightforward task.
Recent works have demonstrated the applicability of CAs to grayscale and RGB (Red, Green and Blue) images when solving processing tasks such as image compression, resizing, skeletonization, erosion/dilation, edge detection, segmentation, forgery detection, content based retrieval and pattern generation (Díaz-Pernil et al. 2014; Dogaru and Dogaru, 2014; Gao and Yang, 2014; Ioannidis et al. 2014; Mardiris and Chatzis, 2014; Minoofam et al. 2014; Rosin and Sun, 2014; Tralic et al. 2014; van Zijl, 2014). However, very little work has been carried out for the application of CAs to processing hyperspectral images (Lee and Bruce, 2010) and, in most cases, the rules are set manually and in an ad-hoc manner.
This chapter describes a methodology for applying CAs to hyperspectral data in order to address different processing tasks. The first one of these tasks is the segmentation of still hyperspectral images, aiming to transform the hyperspectral datacubes into modified versions that are easier to process by, for example, subsequent classification methods. Once this approach has been analyzed and validated, the applicability of CAs to hyperspectral data is pushed to a higher complexity level by introducing the temporal dimension in the processing of sequences of multi-temporal hyperspectral images. In this the case, the cellular automata structures deal with the denoising problem by taking into account the inter-dimensional diversity by jointly processing the spatial, spectral and temporal information of multi-temporal image sequences.
The application of the proposed methodology entails the following advantages:
- •
The CA structures have the capacity of adapting to what the user desires contemplating the fact that there may be multiple ways of solving the processing task.
- •
The transition rules are obtained automatically by following a training procedure based on evolutionary algorithms that avoids having to resort to large training or labelled sets of real images. In the evaluation step of the evolutionary algorithm low dimensional reference samples are used, thus simplifying and accelerating the training process.
- •
The CA structures work with the complete spectral breadth of the images, avoiding projecting the spectral information onto lower dimensional spaces.