| Digital imaging system has been widely used in both military and civilian fields,but the image quality in these applications is often affected by various factors such as the fabrication process of the detectors,the relative motion between the camera and scenes,or the weather condition for imaging.Those characteristics will lead to different kinds of degradations include noise interference,motion-blurring,non-uniformity,or low-resolution.Therefore,image quality improvement is an important pre-processing stage for those visionbased tasks and has become a hotspot for image processing.This thesis analyses some classical image quality improvement subjects such as image enhancement,non-uniformity correction of infrared images,and single image superresolution.Several new algorithms are presented on those topics,the recently proposed new models such as curvature constraints,guided image filter and half quadratic splitting are introduced and applied to improve their performance.Besides,since image quality improvement is always considered as a pre-processing method,which is a time critical task,this paper also designs two GPU parallel acceleration schemes to make sure the process can be carried out real time.Furthermore,this paper also explores CNN based single image reflection removal problem and proposes two novel networks which perform better than the existing state-of-the-art methods.The main achievements of this thesis are summarized as follows.(1)A noise-suppression image enhancement algorithm is proposed.This new method assumes that the gradient values of the smooth image are too weak,thus can be enhanced by a nonlinear transform in the gradient domain.To resolve the noise amplification problem during the gradient transform,a Gaussian curvature filter based scheme is introduced.First,a nonlinear gradient transformation is utilized to enhance the edge and texture of small gradients.Then the enhanced image is reconstructed by optimizing a Euler-Lagrange equation through gradient descent method.Gaussian curvature filter is performed on the input image and its first and second derivatives during each iteration to smooth noise and preserve edges.The experimental results with various images show the proposed algorithm can effectively suppress noises without losing details,which indicates a better performance than traditional gradient transform based algorithm.A GPU parallel implementation is also proposed which runs 200-300 times faster than CPU serial method.(2)The performance of SLP-THP based non-uniformity correction algorithm is seriously affected by the results of SLP filter,which always leads to the losing of image details.To address this problem,an improved SLP-THP based non-uniformity correction method with curvature constraint is proposed.The key point of this method is the new way to estimate SLP component.First,the details and contours of the input image are obtained by minimizing local Gaussian curvature and mean curvature of the image surface respectively.Then,the guided filter is performed to combine these two parts together to get the estimate of SLP component.Finally,this SLP component is taken into THP algorithm to achieve the correction.The performance of the proposed algorithm is verified by several real and simulated infrared image sequences.The experimental results indicate that the proposed algorithm can reduce the non-uniformity without detail losing.After that,a GPU based parallel implementation that runs 150 times faster than C++and 2500 times faster than matlab is presented,which shows that the proposed algorithm has great potential for real time application.(3)Model-based optimization methods have been widely used in varies image restoration solutions and achieved some remarkable results.However,finding out a closed mathematical solution for certain priors remains a great challenge.To resolve this problem,this paper presents an improved model-based algorithm for single image super-resolution.Instead of focusing on specific prior knowledge,we exploit the optimization scheme of general image restoration formula.In our approach,the general format of modeloptimization problem is transformed into an alternant renewal process through half quadratic splitting.This transform can also separate the optimization into a modular structure and allows us to optimize the fidelity term and regularization term separately.Of which the regular optimization process can be considered as a denoising process.Then the guided filter is taken as a denoiser to realize this optimization,which uses local linear transform to keep the detail and L2 norm constraint to smooth the noise of input image.Experiments with benchmark datasets and our own infrared images show that our method can surpass several famous model-based and data-based methods in PSNR and SSIM.(4)Single image reflection removal is of great practical importance for various computer vision tasks.Most non-learning methods try to solve this problem through the model-optimization scheme,which fails to produce promising results due to the shortage of suitable priors to model the difference between the reflection layer and the transmission layer.This paper presents an improved generative adversarial network to resolve this problem.Our new generative network has an encoder-decoder structure with skip-connections,followed by an eight-layer fully convolutional reconstruction module.The attention enhancement block is integrated into each skip-connection of the encoder-decoder module to enhance both channel-wise and spatial-wise feature representation.Furthermore,we also apply the WGAN-GP instead of DCGAN in previous reflection removal networks for adversarial supervision.The WGAN-GP loss,specifically,is combined with L2 pixel loss and L1 VGG19 perceptual loss for training.The experimental results with benchmark datasets indicate that our method outperforms several state-of-the-art networks.(5)A two-stage single image reflection removal network is proposed.Currently,the training set for reflection removal is hard to collect because the refraction of the glass will lead to unaligned transmission and input pairs,as a result,synthesized samples are widely used in various works.For those synthesized images,the transmission and reflection layers are both known,some researches try to make full use of the reflection layer by adding an extra prediction for it and applying a penalty term in the objective function.In our opinion,it is not enough to fully utilize the useful information of the reflection layer,we present a two-stage network to resolve this problem.The first stage of the proposed net is used to predict both transmission and reflection layer,then,instead of just applying penalty terms on these two predictions,we consider the reflection layer as a soft mask and combine these two predictions together with feature maps to feed the second stage,we use the recently proposed gated convolution to refine the transmission layer.Furthermore,we also take the predicted reflection layer as the input of the discriminator which may also contribute some improvements.The performance of all the above-mentioned methods have been validated by rigorous experiments,the results show that the new algorithms in this thesis can surpass some other state-of-the-art methods in terms of image quality improvement.The GPU parallel implementation runs hundreds of times faster than the CPU serial processing method,which indicates great potential for engineering application. |
PDF Full Text Request