Research On Computational Modeling Of Visual Attention And Its Applications

Visual attention is an important characteristic of human visual perception. Humans can easily find the objects of interest in a complex visual scene; however, traditional computer vision can not do well. This paper is primarily concerned with how to develop a human-like computational model of vision attention. Existing bottom-up attention models includes spatial domain models and Fourier transform-based models. Spatial domain models have biological plausibility; but they suffer from high computational complexity. Fourier transform-based models have fast computational speed; but they lack biological plausibility. Following these models, we propose a visual attention model based on the pulsed PCA transform, and furter extend the model to a visual attention model based on the pulsed cosine transform. Moreover, we apply our proposed model of visual attention to ship detection in synthetic aperture radar images and to compressive sensing of image signals. The main innovations of this paper can be described into four aspects as follows:1. A computational model of visual attention based on pulsed PCA is proposed, which exploits the PCA coefficients of images to generate the spatial and motional saliency. The PCA projection vectors can be obtained by the Hebbian learning in neural networks. Moreover, the saliency information in the model can be expressed in brinary codes, which mimic neuronal pulses in the human brain. Therefore, our pulsed PCA model has more biological plausibilities than exsiting Fourier transform-based models. According to the principle of pulsed PCA, we further propose a computational model of visual attention based on the pulsed cosine transform (PCT). The PCT model exploits the DCT coefficients of images to generate visual saliency. The advantage of this model exists in the fact that it need not to estimate the PCA projection vectors, and that DCT has many fast algorithms, which make our proposed PCT model has very fast computational speed and can be applied in a realtime saliency detection system.2. Pulsed PCA model needs to estimate the principal component vectors of natural images. The PCA problem of high-dimensional data (such as images) is difficult to deal with. In order to solve this problem, we propose a fast data-oriented PCA algorithm. In the iteration process, the updated eigenvector is the weighted average of all samples. Thus, it needs not to calculate the covariance matrix of data. This algorithm is capable of overcoming the shortage of the conventional batch PCA approaches, which suffer from high space and time complexity in decomposing the covariance matrix. Moveover, it has faster convergent speed and higher cacluation accuracy than the incremental algorithms.3. Speckles and heterogeneous regions in synthetic aperture radar (SAR) images pose great challenges on automatic detection of ships. This paper proposes a visual attention-based method for ship detection in SAR images. Accoridng to the charisteristics of SAR images, the PCT model is improved and used to generate visual saliency maps of SAR images. Thus, ship signatures in SAR images and their backgrounds can be separated in terms of gray-scale levels, and hence the false alams can be avoided. Our proposed method is very simple, fast, and robust against the speckles and heterogeneous regions in SAR images. It outperforms conventional ship detection approaches.4. A saliency-based compressive sensing method for images is proposed. This method employs the PCT model to generate the the saliency information of the visual scene. According to the obtained saliency information, it allocates more sensing resources to salient regions but fewer to nonsalient regions. Our proposed method takes human visual perception into consideration because human vision would pay more attention to salient regions, but less attention to nonsalient regions. The method improves the reconstructed image quality considerably compared to the case when saliency information is not used.