Research On Technology Of Infrared Image Enhancement Based On Human Visual Model

A passive detection technology based on the thermal radiation of target is used in infrared targets detection system, which has the advantages of self-hiding and all-weather monitoring. As a result, infrared imaging technology is widely used in military detection, civil monitoring and other fields. Infrared images are fuzzy and often characterized by low contrast, blurred texture details due to thermals isotropy radiation and uneven photosensitive response of infrared sensor. Without any processing technology, original infrared images are not suitable for observation. For better observation and monitoring it is necessary to enhance original infrared image. Recently, the enhancement technology, which is based on the characteristics of human visual system, attracts much attention. In order to improve infrared images captured by vehicle mounted equipment in our institute, the visual model, which is constructed under the characteristics of human visual system, has been studied and analysed then research on its application on infrared image enhancements. In this paper, we focus on contrast enhancement, detail enhancement and salient region enhancement. Below are the main works of this paper:1 The principle and characteristics of infrared imaging are been introduced firstly, then the basic features of infrared images are presented. Then the relationship between human visual system and computer vision are introduced. Existing visual models are been reviewed in this paper. Highlighted visual models, which related to this paper, are atmospheric scattering model, Retinex model and visual saliency.2 In computer vision, atmospheric scattering model is the physical model that describes how light propagates in atmosphere. After analyzing atmospheric scattering model, a method based on atmospheric scattering model is proposed to achieve fuzzy phenomenon removing from original infrared images. When calculating the transmission rate, a new method using average filtering is proposed to compute relative depth information. With the help of statistical information in infrared images, the transmission rate is self-adaptive estimated. At last, to solve the low light problem of fuzzy free infrared image, a further method for contrast enhancement and brightness increasing, based on scene complexity computation plateau histogram equalization, is proposed. Experiments show that in both subjective observation and objective quantitative evaluation the performance of the proposed algorithm are pleased, compared to other algorithm in the experiment.3 Infrared images are results of target thermal radiation. When the difference of thermal radiation energy between target and background is huge, restricted by display equipment, it is difficult to show the detail information both in high light regions and dim regions. To deal with this problem, an adaptive detail enhancement method, based on subband-decomposed multi-scale Retinex, is proposed. Firstly, three independent spectrum subbands using subband-decomposed multi-scale Retinex. Then guided image filter is applied to get detail layer and base layer from each subband. Later the basis weight function for detail enhancement is proposed according to characteristic of separate spectrum subband. Adaptive detail enhancement is achieved with basis weight function. Following, a enhanced infrared image is recreated with the base layer and enhanced detail layer. Lastly,in order to eliminate the nonuniformity of gray intensity in the outcome infrared image,a new adaptive way to get gamma curve for gray value remapping is been put forward. Experimental results show that the detail of the enhanced images is upgraded greatly in both high light regions and dim regions. Moreover, the proposed method gets the highest score in objective evaluation.4 Human vision pays more attention to the interesting region than other areas. A method based on salient region detection for layered difference representation of 2D histogram is proposed to achieve visual enhancement. The algorithm detects the salient region by salient filtering and cuts salient region with a threshold for visual perception firstly. Then 2D histogram is calculated for related region in original image of salient region and statistical information in different layer is converted to layer according to the inner relationship of each layer. Following a difference vector is gained though solving a constrained optimization problem of layered difference representation at a specified layer. An original difference vector is defined to represent the character of source image. Output image is reconstructed by a transformation function, which is result of two difference vector for salient region and non-salient region. Experimental results show that the proposed method enhances contrast and details in salient region efficiently while protecting non-salient region in origin image. The objective evaluation parameters in two group experiments illustrate the proposed algorithm get better scores in protecting global mean lighting in non-salient region and increasing PSNR of the whole image compared to other algorithms.At the last of this paper, the summation about the main work achieved and problems need to further study and solve is given.