Researches On Selective Visual Attention

Selective visual attention selection is an important mechanism of Human Visual System (HVS). Our visual system receives an enormous amount of information from the outside world at each moment. But the information conveyed to the high level of brain is actually highly reduced by controlling and tuning of different visual neurons and cells in our brain. This is how attention mechanism works. Due to this mechanism, human beings are able to rapidly attend to the conspicuous locations within the visual field. In the field of computer vision of machine intelligence, the mechanism of visual attention is always taken into account for preprocessing of tasks like target detection, object segmentation, image compression and so on.Bottom-up and top-down mechanisms are considered two main branches of visual attention. The bottom-up mechanism is totally driven by stimuli and data, which is deeply studied in our thesis. Models categorized to this type include Neuromorphic Vision Toolkit (NVT) which is completely inspired by biological discoveries, Phase Quaternion Fourier Transform (PQFT) which utilizes quaternion to implement phase spectrum transform and Frequency Tuned Saliency (FTS) based on engineering manipulations. Among the three models, NVT and PQFT turn out to be merely effective on saliency prediction of small size objects or edge information. On the other hand, FTS mainly tackles the problem of saliency region detection of large objects.First of all, considering pros and cons of existing models, we propose a frequency band selection model from the perspective of frequency domain, which is based on the discovery of non-Classical Receptive Field (nCRF). The model analyzes the information related to both low and high spatial frequency. Each frequency band is measured in terms of saliency and the optimal one(s) is/are picked out to build a saliency map. Our model is capable of predicting saliency of objects with different sizes and outperforms most models which only deal with certain types of data.After that, we introduce the whitening method to improve our band selection attention model. For one thing, the whitening processing, which decorrelates data samples, is utilized to effectively extract saliency information of feature maps of diverse scales. This idea helps to avert the high computation cost of traditional segmentation or patching approach for saliency extraction. For another, based on the improvement of evaluation metric, we modify the weight function for band selection. The improved model is proved to be superior on fixation datasets in terms of new evaluation metric and also has good performance on segmentation datasets. What’s more, the model is able to deal with saliency prediction of most cases of psychological patterns.At last, we propose a framework which employs our band selection model as bottom-up factor and line feature as top-down factor for airport detection in remote sensing images. Common methods for detection are always based on sliding window or segmentation which incurs expensive computation cost. However, the mechanism of visual attention can rapidly locate region of interest (ROI) with low cost. The experiment results show the detection rate and false alarm control could be improved by combining the bottom-up and top-down attention mechanisms. Beyond that, we also make analysis on the influence of different fusion strategy, amount of training samples and number of ROIs.