基于Transformer和双残差网络的图像去模糊算法研究

在单幅图像去模糊网络的设计中,已广泛采用了由“粗到精”的策略。经典的基于深度学习的多尺度网络模型通常将子网络与多尺度输入图像堆叠,并从底层子网络到顶层子网络逐渐提高图像的清晰度,将不可避免地产生图像特征信息丢失问题以及较高的时间成本。针对上述问题,构建了一种基于编解码器结构的多输入多输出U型网络。首先,在编码器结构中采用Transformer模型代替传统卷积网络模型以充分获取图像特征信息;其次为了恢复高频细节特征,在解码器结构中设计了双残差网络;最后,提出多尺度融合算法,有效融合了多尺度特征。在GoPro上的仿真实验表明,通过引入Transformer,所提算法分别在PSNR与SSIM评价指标和运行时间上均优于主流算法,PSNR与SSIM值分别为31.15 dB、0.957 7,同时从视觉上可以得到更多的图像细节。     

基于增强多尺度特征网络的图像去模糊

现有的多尺度多阶段图像去模糊方法存在多尺度图像特征提取利用不充分和阶段加深导致特征信息丢失的问题。针对上述问题,提出了一种基于增强多尺度特征网络的图像去模糊方法。首先,提出一个多尺度残差特征提取模块,在两个分支中使用不同大小的卷积核扩大感受野,充分提取不同分辨率大小图像的特征信息。其次,提出跨阶段注意力模块,过滤和传递图像的关键特征。最后,设计类似于跳跃连接的跨阶段特征融合模块,弥补特征损失,同时融合来自不同大小输入图像的特征信息,丰富空间特征信息,以更好地处理纹理细节。在GoPro和HIDE数据集上的实验结果表明,所提方法能够很好地重建图像。     

基于多尺度与残差网络的图像超分辨率重建

针对目前图像超分辨率重建算法在提取图像信息时过于单一地累加卷积层导致的深层网络梯度消失和图像信息丢失的问题,提出了一种基于多尺度与残差网络的图像超分辨率重建算法。该算法使用多尺度密集连接的卷积核代替单一累加的卷积核,充分提取输入低分辨率图像信息并实现通道特征维度的复用;利用残差网络,多层次地对丢失图像信息进行补充并对深层网络模型的梯度问题实现了抑制,在反向传播的过程中帮助全网络模型自适应地完成对权重更新;最后以非线性映射的方式输出最终重建图像。实验表明,所提算法在测试集上的峰值信噪比和结构相似性与对比算法相比均有所提升;在与目前主流算法对比中获得了细节信息更加丰富、边缘纹理更加清晰的重建图像。     

基于多尺度优化和动态特征融合的图像去模糊研究

目前采用U-Net结构的去模糊算法存在细节损失、图像质量欠佳等问题,因此对U-Net进行改进,提出一种基于多尺度优化和动态特征融合的图像去模糊方法。首先针对细节损失,提出一种精简且有效的多尺度残差注意力模块(Multi-Scale Residual Module, MSRM),通过增加特征尺度多样性来提取更精细的图像特征。此外,为了将更有利的特征传递到解码部分,在跳跃连接处设计动态特征融合模块(Dynamic Feature Fusion Module, DFFM),采用注意力加权的方式选择性融合不同阶段的编码特征。该算法采用多尺度内容损失和多尺度高频信息损失进行约束训练。在GoPro和RealBlur数据集上的实验结果表明,这种方法能有效改善图像质量,复原更丰富的细节信息。与现有去模糊算法相比,本文算法在主观视觉和客观评价等方面均具有一定优势。     

多尺度卷积神经网络的噪声模糊图像盲复原

图像盲复原是从一幅观测的模糊图像恢复出模糊核和清晰图像,传统盲去卷积算法采用简化模型估计模糊核,导致预测模糊核与真实值误差较大,最终复原结果不理想。针对此问题提出一种基于改进残差模块的多尺度卷积神经网络模型,采用端到端模式,无需估计模糊核。提出了一种基于限制网络输入的改进Wasserstein GAN （WGAN）,增加了一层输入限制层,能够限定参数初始值,提高了网络收敛速度。设计了多重损失函数,融合了基于多尺度网络的感知损失和基于条件式生成对抗网络的对抗损失。实验结果表明:所提方法在定量和定性评价指标上优于已有的代表性方法,并且运行速度比相近算法快了4倍。     

基于多尺度近端特征拼接网络的高光谱图像分类方法

针对基于传统卷积神经网络模型的高光谱图像分类算法细节表现力不强及网络结构过于复杂的问题,设计了一种基于多尺度近端特征拼接网络的高光谱图像分类方法.通过引入多尺度滤波器和空洞卷积,在保持模型轻量化的同时可以获取更丰富的空间-光谱判别特征,并提出利用卷积神经网络近端特征间的相互联系进一步增强细节表现力.在3个基准高光谱图像...     

基于有理数多项式先验模型的图像盲去模糊

图像盲去模糊问题是当今图像处理领域的热点问题之一.基于混合高斯先验模型的变分贝叶斯去模糊算法可以有效地复原模糊图像,成为一种重要的图像去模糊算法.虽然混合高斯先验模型可以很好地逼近自然图像的梯度分布,但是该模型在图像梯度值较大处往往会产生过拟合导致去模糊后的图像产生振铃效应,严重影响了图像可读性.利用有理数多项式先验模型代替混合高斯模型逼近自然图像的梯度分布,克服算法的上述缺点.有理数多项式函数的分母多项式强制函数在梯度值较大值时平滑,所以有效地避免了过拟合现象的发生,从而使得模糊核估计得更准确,减少振铃效应.实验结果表明了算法的可行性和有效性.     

双分支特征提取与循环细化的动态场景去模糊

针对现有的动态场景图像去模糊方法存在的特征提取不准确、未充分利用有效特征的问题,本文提出了一种基于双分支特征提取与循环细化的动态场景图像去模糊网络。整个网络包括特征提取网络、循环细化网络(cyclic refinement network, CRN)、图像重建(image reconstruction, IR)3部分。其中,特征提取网络包括模糊图像细节和轮廓特征(contour feature, CF)的提取,以残差单元作为特征提取网络的基本单元;循环细化网络通过交替融合轮廓特征和细节特征(detail feature, DF)来细化特征图,得到模糊图像的细化特征(refinement feature, RF);最后,在图像重建阶段,复用轮廓和细节特征,结合残差学习策略将轮廓特征、细节特征和细化后的特征逐级融合后通过非线性映射的方式重建清晰图像。在广泛使用的动态场景模糊数据集GOPRO上的实验结果表明,该方法的平均峰值信噪比(peak signal to noise ratio,PSNR)达到31.86,平均结构相似度(structure similarity,SSIM)达到0.947...     

Improved image deblurring based on salient-region segmentation

Image deblurring techniques play important roles in many image processing applications. As the blur varies spatially across the image plane, it calls for robust and effective methods to deal with the spatially-variant blur problem. In this paper, a Saliency-based Deblurring (SD) approach is proposed based on the saliency detection for salient-region segmentation and a corresponding compensate method for image deblurring. We also propose a PDE-based deblurring method which introduces an anisotropic Partial Differential Equation (PDE) model for latent image prediction and employs an adaptive optimization model in the kernel estimation and deconvolution steps. Experimental results demonstrate the effectiveness of the proposed algorithm.     

Multi-stage feature-fusion dense network for motion deblurring

Although convolutional neural networks (CNNs) have recently shown considerable progress in motion deblurring, most existing methods that adopt multi-scale input schemes are still challenging in accurately restoring the heavily-blurred regions in blurry images. Several recent methods aim to further improve the deblurring effect using larger and more complex models, but these methods inevitably result in huge computing costs. To address the performance-complexity trade-off, we propose a multi-stage feature-fusion dense network (MFFDNet) for motion deblurring. Each sub-network of our MFFDNet has the similar structure and the same scale of input. Meanwhile, we propose a feature-fusion dense connection structure to reuse the extracted features, thereby improving the deblurring effect. Moreover, instead of using the multi-scale loss function, we only calculate the loss function at the output of the last stage since the input scale of our sub-network is invariant. Experimental results show that MFFDNet maintains a relatively small computing cost while outperforming state-of-the-art motion-deblurring methods. The source code is publicly available at: https://github.com/CaiGuoHS/MFFDNet_release.     

基于边缘信息的运动模糊图像的鲁棒盲复原

为获得适用于不同模糊图像且简捷的图像盲复原方法,提出了一种稳健的从单幅模糊图像中求取模糊核并对图像去模糊的图像盲复原方法。根据模糊图像与非模糊图之间的边缘关系求模糊核,并在多尺度框架下针对各个子算法设定自适应参数,从而构建一个稳健的图像盲复原方法。对复原结果用4种无参考的图像质量评价方法的评价结果显示,本文方法在噪声和...     

Robust blind motion deblurring using near-infrared flash image

In light-limited situations, camera motion blur is one of the prime causes for poor image quality. Recovering the blur kernel and latent image from the blurred observation is an inherently ill-posed problem. In this paper, we introduce a hand-held multispectral camera to capture a pair of blurred image and Near-InfraRed (NIR) flash image simultaneously and analyze the correlation between the pair of images. To utilize the high-frequency details of the scene captured by the NIR-flash image, we exploit the NIR gradient constraint as a new type of image regularization, and integrate it into a Maximum-A-Posteriori (MAP) problem to iteratively perform the kernel estimation and image restoration. We demonstrate our method on the synthetic and real images with both spatially invariant and spatially varying blur. The experiments strongly support the effectiveness of our method to provide both accurate kernel estimation and superior latent image with more details and fewer ringing artifacts.     

基于多尺度多路径FCN的视网膜血管分割

视网膜血管的形态结构信息可以为糖尿病、高血压等疾病提供诊断依据。提出了一种基于多尺度多路径的全卷积神经网络的视网膜血管分割方法。首先,利用空洞卷积代替池化层和上采样操作,在不增加参数的情况下增加感受野,避免了细节信息的丢失;其次,通过使用不同空洞率的空洞卷积实现图像数据的多尺度特征提取,充分学习图像的多尺度特征,避免网络过深,并提升了细小血管的提取能力;同时,利用跳层结构在网络中建立多条信息流通路径,通过多路径信息流充分传递多尺度特征信息,提高网络预测效果。实验结果表明,该算法在DRIVE数据集上的平均准确度、灵敏度和特异性分别为95. 46%、81. 24%、97. 77%,取得了较好的视网膜血管的分割效果。     

Hyperspectral image quality based on convolutional network of multi-scale depth

Hyperspectral imagery has been widely used in military and civilian research fields such as crop yield estimation, mineral exploration, and military target detection. However, for the limited imaging equipment and the complex imaging environment of hyperspectral images, the spatial resolution of hyperspectral images is still relatively low, which limits the application of hyperspectral images. So, studying the data characteristics of hyperspectral images deeply and improving the spatial resolution of hyperspectral images is an important prerequisite for accurate interpretation and wide application of hyperspectral images. The purpose of this paper is to deal with super-resolution of the hyperspectral image quickly and accurately, and maintain the spectral characteristics of the hyperspectral image, makes the spectral separability of the substrate in the original image remains unchanged after super-resolution processing. This paper first learns the mapping relationship between the spectral difference of low-resolution hyperspectral image and the spectral difference of the corresponding high-resolution hyperspectral image based on multiple scale convolutional neural network, Thus, apply this mapping relationship to the input low-resolution hyperspectral image generally, getting the corresponding high resolution spectral difference. Constrained space by using the image of reconstructed spectral difference, this requires the low-resolution hyperspectral image generated by the reconstructed image is to be close to the input low-resolution hyperspectral image in space, so that the whole process becomes a closed circulation system where the low-resolution hyperspectral image generation of high-resolution hyperspectral images, then back to low-resolution hyperspectral images. This innovative design further enhances the super-resolution performance of the algorithm. The experimental results show that the hyperspectral image super-resolution method based on convolutional neural network improves the input image spatial information, and the super-resolution performance of the model is above 90%, which can maintain the spectral information well.     

多尺度残差特征融合的轻量级真实图像超分辨率重建

基于深度学习的真实图像超分辨率(super-resolution, SR)重建算法目前存在参数量过大的问题,为解决该问题,提出了一种多尺度残差特征融合的轻量级真实图像SR重建算法。首先利用深度可分离卷积和复用卷积针对多尺度特征提取块进行改进,在提取特多尺度特征的同时实现了模块的轻量化,参数量仅为改进前的7.5%。其次使用残差特征融合操作将4个多尺度深度可分离特征提取块(multi-scale depthwise separable block, MSDSB)聚合成一个残差特征融合块,以减少残差路径长度。然后使用增强型注意力模块从通道和空间维度进行自适应调整以提升算法性能。最后使用自适应上采样模块获得SR重建图像。在消融实验中,本文算法重建性能超过原始算法,且参数量仅为3.53×10⁶,是原始算法的34.5%。在对比实验中,其重建性能超过了当前主流算法,与组件分而治之(component divide-and-conquer, CDC)算法相比,PSNR和SSIM指标分别提升了0.01 dB与0.001 0,且参数量仅为组件CDC算法的8.84%,在保证重建性能的同...