Research On Infrared Image Enhancement And Hardware Acceleration Technology

Infrared(IR)imaging equipment is widely used in areas of military defense,security monitoring,medical diagnosis and many other fields because of its advantages on night vision,anti-interference,high precision,and good concealment.However,with the limitations of IR sensors,IR images have disadvantages such as low signal-to-noise ratio,low contrast,and blurred details.Therefore,the IR image enhancement technology that provides a solution to the disadvantages above has become a research hotspot in IR imaging.Nowadays research on IR image enhancement technology is focused on two directions: HDR details enhancement and multi-source merge.It is found that the processing results of the existing detail enhancement algorithms have problems including gradient reversal,noise amplification and low contrast while the fusion enhancement algorithm faces low contrast,blocking effect,and loss of small targets.In order to solve these problems,this thesis has done a deep research based on the two approaches of which the universal model is presented.The model is total variation based intensity fidelity and gradient fidelity optimization model.This thesis applies the minimization model with gradient fidelity term and intensity fidelity term in Dynamic Range Compression and Detail Enhancement(DRC&DE)technology for High Dynamic Range(HDR)IR image.A new DRC&DE algorithm is then proposed of which the target of strength constraint is the HDR image processed by histogram equalization with a threshold,effectively compressing the dynamic range of the raw HDR image and improving the result.Gradient constraint target is acquired by adaptive attenuation of raw image's gradient,which enhances the details effectively and avoids the effect of gradient reversal and noise amplification that none of the methods available has solved simultaneously.The minimization model with gradient fidelity term and intensity fidelity term is applied in the fusion technology for fusing IR image and Visible(VIS)image,and then a total variation image fusion model based IR image enhancement algorithm is proposed in this thesis.It solves three major problems of target loss,low contrast,and blocking effect which is common in similar algorithms.The intensity fidelity term uses IR images to constrain the intensity of the fusion image so that the characteristics of the IR target of fused image is obvious enough.The gradient fidelity term uses the fusion gradient of the VIS image gradient and the IR image gradient to constrain the gradient of the fusion image.In this way the problem of loss of IR targets that are prone to similar algorithms is effectively solved.In this thesis,the structure tensor adaptive weighted fusion of the source image's gradient is adopted to improve the adaptivity of the algorithm.This thesis proposes a multi-scale gradient optimization scheme based on the pyramid model,which can improve the speed and accuracy of gradient relaying in iterations,thereby improving the fusion efficiency and effectiveness.Compared with the classical image enhancement method and the superb fusion algorithm,the enhanced image in this paper has the advantages of clear infrared target,rich details,and strong overall contrast in subjective evaluation.It has a higher level in objective evaluation criteria.In order to improve the processing efficiency of the algorithm,this paper studies the low-power and high-performance CPU+FPGA heterogeneous computing technology.The proposed fusion-based image enhancement algorithm is implemented on a heterogeneous computing platform.This paper balances the data transmission time and computing time by optimizing kernel,reasonable memory partitioning,kernel vectorization,and loop expansion.This improves the efficiency of the algorithm.Compared with the traditional serial computing architecture,the CPU+FPGA heterogeneous platform is capable of processing images 4.6 times faster speedup when the image is enhanced with a resolution of 1024*1024 and at the same time reducing the computational energy consumption by 70%.This research project can provide low-power real-time solutions for image processing algorithms with high computational complexity and high degree of parallelism.