Study On Image Visual Saliency Analysis And Its Applications

The human visual system can find and pay attention to the interested content quickly from the images or scenes. That how to imitate the visual cognitive mechanism to establish the computational model of visual saliency, realizing the detection and recognition of objects is the current research focus in the field of computer vision.This thesis does further research and improvement on the bottom-up visual saliency model according to the visual attention mechanism, computational theory of vision and the computational models of visual saliency, as well as its applications on static images and video image sequences.It sums up the general framework model which can be divided into three parts of the feature detection and description, the saliency map generation and the saliency map fusion after the review and analysis of the bottom-up visual saliency model, as well as researches and analyses the different processing method used by the parts of the framework, including it’s scope of application and problems. Aiming at the lack of significant feature selection and the complicated algorithm of Itti model, the significant improved calculation method, that is combined the significant global color and the significant direction and carries on multi channels decomposition in the Lab space which is more in line with the visual characteristics, using Gaussian differential filtering for each channel color feature extraction to do the global mean color saliency map computation and the direction selectivity of Gabor filter to gain the saliency map extraction in different directions, is put forward. Furthermore, it turns out that this method in the thesis in the integrity of the salient object extraction and the similar degree of artificial labeled graph is superior to the results of the compared model when the improved algorithm model, Itti model and SR model were experimented on the MSRA visual saliency detection database.It also studies the sea pearl target recognition method based on multi feature fusion and pattern recognition of color, texture and shape and designs and develops seawater pearl recognition prototype system using of Matlab and SQL Server. Then, the significant calculation model is applied to the target pearl image in the system of automatic detection. And taking advantage of the the principle of saliency detection of SR model, this paper presents a extraction method of static background of multi channel video frames with residual relationship between spectrum calculation. The method, according to the fixed time interval from the video image sequence, carries on the statistics in the log spectral domain multi color channel, filters significant target information in multi frame equalization, and reconstructs the background image after inverse transform. In addition, the experiment verifies the validity of the method. Besides, combined with the relationship between adjacent frames in video sequences, the Otsu algorithm is improved by using the genetic simulated annealing algorithm, binding the three frame difference method to achieve the detection of moving target. And the experiment proves that the method makes it possible to improve the accuracy of the two values of the threshold and the moving target detection more complete.