Research On Improved Saliency Detection Model In Frequency Domain And Its Application

Visual attention mechanism can break through the bottleneck of information processing,and it only allows some valuable input perception information as the source of visual consciousness.Similarly,giving priority to the saliency objects when it allocates the computing resources in practical application can greatly contributes to improve the computer processing speed.Therefore,it is very meaningful for researching on Visual Attention Model(VAM)in the field of computer vision and image processing.Generally speaking,despite VAM in spatial domain can get good results,but its computing time complexity is usually high,and there are too many manual adjustable parameters etc.Therefore,visual attention models in frequency-domain are introduced.However,VAMs in frequency domain tend to detect edges instead of the interior of the saliency objects and are short of biological plausibility.Therefore,This thesis builds a saliency detection model based on spatio-frequency information,and applies it to image classification.The main contribution of this thesis are as follows:(1)VAMs in frequency domain usually highlight the edge of saliency objects.This thesis improves the HFT model in two aspects,and proposes IHFT(Improved HFT)model.Firstly,the coefficients of the input hyper complex image are well designed to be more in line with the human visual characteristics.Secondly,the spatial contrast function and density distribution function are employed to select the optimal saliency map.Experimental results show that the average AUC value of the IHFT model is 0.8454,and F–measure achieves to 0.8300,which is better than related models.(2)In order to be more in line with human visual perception,the depth information is introduced in hyper complex representation of the image in HFT model.For IHFT model still can not even fully highlight the saliency objects,HC(Histogram-based Contrast)model is employed to improve the performance of IHFT model.IHFT model can better detect saliency of images with complex texture background,but all tend to detect edges instead of the interior of the saliency objects;HC model can better detect the interior of saliency objects,however HC model cannot detect saliency of images with complex background.This thesis tries to fuse the IHFT and Histogram Based Contrast(HC)together non-linearly to obtain the final saliency map.Experimental results show that IHFT + HC model can not only highlight the saliency objects evenly,but can deal with complicated background images,and the average AUC value of the model is 0.8985,and F-measure achieves to 0.8414,that is to say,the proposed model performs favorably against stated-of-the arts.(3)This thesis proposes a image classification method based on VAM First of all,to reduce the computational complexity,image classification is taken on the premise of extracting the saliency region.Secondly,texture features and PCNN time signatures based on saliency region greatly reduce the feature dimension,more importantly,the two features can express the essential content of the image more reasonably.Finally,image multi classification is proposed.Experimental results show that the image classification precision rate of the proposed method based on visual attention model is 94.42%,which increases 5.04% compared with the method based on the whole image.