Research On Infrared Images Saliency Detection Methods

As digital image information keeps increasing rapidly in recent years,computers are expected to obtain the ability to focus on only several important areas of digital image information,just like the human visual attention mechanism.In that case,people can deal with massive image information quickly while using a limited amount of computer resource.Though a large number of researchers have devoted in the field of image saliency detection,there is still a shortage of saliency detection methods based on infrared image characteristics.However,owing to the advantages of being hardly affected by lights and its significant ability to penetrate smokes,infrared images saliency detection methods have a great potential value in both research and application.To solve this problem,on the basis of current researches on images saliency detection,this thesis analyzes the advantages and disadvantages of the existing image saliency detection methods,and proposes two infrared image saliency detection methods combined with the characteristics of infrared images.Results show that the detection methods proposed in this thesis can achieve better performances when compared with other image saliency detection methods.The main contents of this thesis are as follows:Considering the disadvantages in existing saliency detection methods,this thesis proposes an infrared image saliency detection method combined with human visual theory and spectral residual,and has the ability to detect infrared saliency objects efficiently.First this method divide the infrared images into several regions by using the graph-based segmentation method.Next,on the basis of the contrast theory of human visual research,the proposed algorithm calculates the contrast information between one region and other regions by using the histogram to accelerate the calculation.Moreover,this thesis takes into account the distance information between two regions and size information of regions.Finally,according to the principle of information theory,this thesis transforms images into frequency domain,and removes the remaining spectrum parts so that the redundant parts of the image are restrained.Through combining the two saliency maps,the method of this thesis can detect the saliency object in infrared image more accurately when compared with existing detection methods.As for the limitations that some complex structural features within the image may not be described by artificial features accurately,this thesis proposes an infrared image saliency detection method based on convolution neural network.This thesis first uses the graphbased segmentation method to segment the image into different scales,and then applies the two-channel convolution neural network from two different levels for feature extraction.Next the method output image features through two fully connected layers,and the significant images at each scale are calculated.Finally,the significant images at multiple scales are merged.The results of the comparison experiments show that the proposed method is more accurate when detecting the saliency object in infrared images.