9.3 Conclusions

The major parts of this book have introduced visual attention models (saliency map models) which consider both the biological basis and the feasibility for implementation with computers, and their results are represented as saliency maps. The presented models are different from pure computer vision modelling and pure biological level modelling, and this is where the biological facts meet engineering applications. In other words, the emphasis of this book lies in the intersection area between biology and engineering. It is worth noting that different models have advantages and disadvantages. At the current stage of technological development in these areas, one model may have good performance for some images but fail in other images, and another model probably exhibits the opposite. For different databases of natural images, the performance of the same model may be different.

Pure bottom-up models in the spatial domain are based more on biological facts and statistical signal processing rules; those in the frequency domain use similar biological rules (such as whitening, eccentricity of visual acuity on the retina, division normalization, etc.), but operate in the frequency domain, which greatly reduces the computational cost.

The models combining top-down information are probably multifarious because the working of knowledge storage and modulation in the high level cortex is not clear yet. Diversified implementation of the top-down mechanism represents the different conjectures.

The scope of applications of saliency map models is wide due to their fundamental nature. Computer vision and image processing are the major ones among the range of possible applications. Typical application examples have been presented as case studies in the book.

We have discussed open research issues in the previous section. In fact, apart from the aforementioned nine topics (in Section 9.2), there are many more issues, such as the causation of change blindness, the consistency between different kinds of computational models, the relationship between feature binding and the binding of time synchronization and so on. Future biological research progress and new findings on biology and psychology (as well as other related areas) will promote computational modelling improvements and boost the applications, since visual attention plays an important role for many vision-related engineering problems.