Research Of Visual Saliency Detection Algorithm Based On Hypercomplex Transform Domain

With the popularization of electronic products and the development of network transmission,digital information has been increased rapidly,such as image and video.We can be in touch with a lot of images and videos in our daily life and work.Facing the massive digital information,how to select the valuable information quickly and efficiently is a very meaningful research subject.Saliency detection algorithm can detect salient information from image and video,which can be helpful for the follow-up of many visual tasks.Saliency detection algorithm based on frequency domain has fast characteristics,and using hypercomplex to represtent color images can reflect the color features of the image very well.Therefore,it has a very important theoretical meaning and practical value to study a saliency detection algorithm based on hypercomplex transform domain.In this thesis,the main research content includes the following three aspects*:1)In order to improve the integrity of the salient regions detected by the frequency domain saliency detection algorithm,a saliency detection algorithm based on Hypercomplex Fourier transform(HFT)amplitude spectrum optimization is proposed.When we analyze the amplitude spectrum of HFT,we find that the higher amplitude parts of the entire frequency spectrum are caused by the background information of the image,and the local high amplitude parts of the frequency spectrum are caused by the salient regions of the image,so we only need to suppress the higher amplitude parts to highlight the salient information.Firstly,the image is transformed to frequency domain by HFT.Then five appropriate thresholds are selected according to the experiment,and the amplitude spectrum above the threshold are smoothed.Finally,the raw saliency maps are calculated by inversing transform with combining the smoothed amplitude spectrum and the original phase spectrum and the raw saliency map with the smallest value of entropy as the final saliency map.The comparision experiments show that the proposed algorithm has a higher saliency detection precision.2)In order to improve the boundary sharpness of the salient regions detected by the frequency domain saliency detection algorithm,a saliency detection algorithm based on Hypercomplex Discrete Cosine Transform(HDCT)priori boundary is proposed.HDCT has the characteristics of energy concentration at low frequency,so this thesis uses the low frequency energy of HDCT to represent the feature of image block.And the saliency map is calculated based on the different in energy characteristics between the image block and the boundary block.Firstly,we transform the image block to frequency domain by HDCT.Then we use the low frequency energy in the frequency spectrum to represent the feature of the image block.Finally,we use the prior boundary to compute the salient value of the image block,and normalize the salient value of the image block to calculate the saliency map.The comparision experiments show that the algorithm has a high detection precision.3)In order to improve the detection accuracy of the visual fixtion by the frequency domain saliency detection algorithm,a visual fixtion detection algorithm based on Hypercomplex Wavelet Transform(HWT)depth perception is proposed.The HWT transform of the image can generate multiple sub-bands of detail maps in multiple directions,which can reflect the details of the image more comprehensively.Using the convolution network to train the detail map can better extract the features of visual fixtion.Firstly,the image is decomposed by first-level HWT,then the redundant network is used to fuse the multilayer detail sub-bands maps into feature maps with a lower dimension.Finally,an improved VGG16 model is used to detection the visual fixation.The comparison experiments show that the proposed algorithm improves the detection accuracy of fixation.