基于非局部稀疏的圖像去噪與平滑方法研究
發(fā)布時間：2018-09-10 19:19
【摘要】：計算機視覺的研究目標(biāo)是通過圖像或者視頻理解場景,圖像處理技術(shù)是實現(xiàn)該目標(biāo)的關(guān)鍵技術(shù)。圖像去噪和平滑是圖像處理領(lǐng)域的基礎(chǔ)問題。在現(xiàn)實生活中,由于傳感器受環(huán)境影響、傳輸信道被干擾等原因,在圖像獲取與傳輸?shù)倪^程中,噪聲被不可避免的引入,從而造成圖像失真。圖像失真現(xiàn)象必然會對圖像特征提取、場景理解等后續(xù)工作造成干擾,從而影響計算機智能處理任務(wù)的準(zhǔn)確性。圖像去噪的研究目標(biāo)是將噪聲從有噪圖像中分離出來,更好地還原圖像的真實信息,特別是邊緣和紋理細節(jié)。如果能夠?qū)Ω蓛魣D像建立較好的表達模型,就可以保留更多的有用信息,保證復(fù)原圖像的準(zhǔn)確性。圖像平滑的目標(biāo)則是通過分離圖像的結(jié)構(gòu)與細節(jié),用來提取對于人類視覺感知最重要的結(jié)構(gòu)邊緣,能夠為更高級的計算機視覺任務(wù)打下堅實的基礎(chǔ)。圖像去噪和圖像平滑作為基礎(chǔ)的圖像處理技術(shù),在空間項目、醫(yī)學(xué)、考古學(xué)、工業(yè)機器視覺、軍事識別、衛(wèi)星圖像處理等領(lǐng)域有著廣泛的應(yīng)用。但是由于圖像內(nèi)容復(fù)雜多變,圖像去噪和圖像平滑的研究面臨一系列問題和挑戰(zhàn)。首先,依靠現(xiàn)有數(shù)學(xué)工具還無法準(zhǔn)確地描述圖像,現(xiàn)有方法基于各種假設(shè)建立圖像的表達模型,存在一定的局限性；其次,人類視覺系統(tǒng)的工作機制非常復(fù)雜,目前對于人類感知原理的研究還處于初級階段,因此無法以明確的數(shù)學(xué)模型定義一幅圖像中具有視覺意義的特征。本文圍繞圖像去噪和平滑問題的研究熱點和難點展開研究,提出基于非局部稀疏的圖像處理方法。新方法能夠充分利用自然圖像本身的有用信息,有效彌補了數(shù)學(xué)模型的缺陷,以數(shù)據(jù)驅(qū)動的方式大大提高了圖像去噪和平滑的效果。主要工作包括：1.提出了基于PCA字典的自適應(yīng)稀疏編碼去噪方法。通過分析PCA字典上稀疏編碼誤差的統(tǒng)計特性,采用拉普拉斯函數(shù)近似編碼誤差的分布,基于后驗估計理論提出一個新的非局部稀疏編碼模型。新模型中用于平衡保真項與非局部約束項的正則化參數(shù)是自適應(yīng)確定的。為獲得可靠的稀疏編碼估計,提出了基于濾波的迭代收縮算法。濾波可以有效抑制后向投影過程的噪聲,進一步得到稀疏編碼的魯棒估計。新方法有效提高了編碼準(zhǔn)確率,從而取得很好的紋理保留和噪聲去除效果。2.提出了基于低秩和梯度稀疏的圖像平滑方法。通過對自然圖像結(jié)構(gòu)和紋理特征的分析,基于自然圖像非局部自相似性提出一個圖像塊分組低秩先驗,然后結(jié)合平滑圖像的全局梯度稀疏先驗提出一種新的圖像平滑優(yōu)化方法。低秩先驗項約束了平滑圖像中相似分組內(nèi)部各圖像塊的強相關(guān)性,可以去除小尺度噪點和細節(jié)、保持細長結(jié)構(gòu)邊緣,保證一致的平滑效果。針對新的目標(biāo)能量函數(shù)的優(yōu)化問題,給出了基于交替迭代近似求解算法的詳細流程。新方法能夠達到一致性較高的平滑效果,在去除細節(jié)的同時保持重要的結(jié)構(gòu)邊緣。3.提出了非局部梯度聚集圖像平滑方法。通過分析自然圖像梯度圖的特點,基于非局部自相似性提出平滑圖像梯度圖的非局部聚集約束項,將該約束與梯度L0范數(shù)最小化先驗結(jié)合得到一個新的優(yōu)化框架。然后給出了交替迭代算法用于高效求解新能量模型的優(yōu)化問題。新方法的非局部約束以數(shù)據(jù)驅(qū)動的方式削弱了相似塊之間梯度的不一致性,能夠有效地去除復(fù)雜區(qū)域的細節(jié)、保持對比度不明顯的有意義結(jié)構(gòu)。與現(xiàn)有方法相比,新方法的平滑結(jié)果不僅一致性高,而且能夠保持結(jié)構(gòu)邊緣不移位。4.研究了圖像平滑在智能圖像處理中的應(yīng)用。平滑圖像在去除瑣碎細節(jié)的同時保留了對于人類視覺系統(tǒng)非常關(guān)鍵的結(jié)構(gòu)信息,對于內(nèi)容相關(guān)的圖像處理問題具有很強的應(yīng)用價值。首先研究了圖像平滑在圖像放縮、圖像編輯等問題中的應(yīng)用。然后基于平滑方法提出一種新的多尺度空間構(gòu)造方法,并研究了多尺度空間在顯著性檢測問題中的應(yīng)用。實驗結(jié)果顯示圖像平滑本身和多尺度空間在各類應(yīng)用中都起到了很好的提升作用。
[Abstract]:The research goal of computer vision is to understand the scene through image or video. Image processing technology is the key technology to achieve this goal. Image denoising and smoothing are the basic problems in the field of image processing. Noise is unavoidably introduced into the image, resulting in image distortion. Image distortion will inevitably interfere with the follow-up work of image feature extraction and scene understanding, thus affecting the accuracy of computer intelligent processing tasks. Information, especially edges and texture details. If a good representation model can be established for a clean image, more useful information can be retained to ensure the accuracy of the restored image. Advanced computer vision tasks lay a solid foundation. Image denoising and image smoothing as the basis of image processing technology, in space projects, medicine, archaeology, industrial machine vision, military recognition, satellite image processing and other fields have a wide range of applications. There are a series of problems and challenges. Firstly, relying on the existing mathematical tools can not accurately describe the image, the existing methods based on various assumptions to establish the image representation model, there are certain limitations; secondly, the human visual system working mechanism is very complex, the current research on the principle of human perception is still in its infancy. It is impossible to define visually meaningful features in an image by a definite mathematical model. This paper focuses on the research hotspots and difficulties of image denoising and smoothing, and proposes an image processing method based on non-local sparseness. The main work includes: 1. An adaptive sparse coding denoising method based on PCA dictionary is proposed. By analyzing the statistical characteristics of sparse coding errors in PCA dictionary, Laplace function is used to approximate the distribution of coding errors and a posterior estimation theory is proposed. A new non-local sparse coding model is proposed in which the regularization parameters for balancing fidelity terms and non-local constraints are determined adaptively. To obtain reliable sparse coding estimation, an iterative shrinkage algorithm based on filtering is proposed. Filtering can effectively suppress the noise in the backward projection process and further obtain sparse coding. Robust estimation. The new method effectively improves the coding accuracy and achieves good texture preservation and noise removal. 2. An image smoothing method based on low rank sum gradient sparseness is proposed. Then a new image smoothing optimization method is proposed based on the global gradient sparse priori of smoothing image. The low rank priori constrains the strong correlation of each image block in the similar grouping of smoothing image. It can remove the small scale noise and details, maintain the edge of slender structure and ensure the consistent smoothing effect. A detailed flow chart based on alternating iteration approximation algorithm is given to solve the optimization problem. The new method can achieve high uniform smoothing effect and maintain important structural edges while removing details. 3. A non-local gradient clustering image smoothing method is proposed. The characteristics of natural image gradient map are analyzed and non-local self-similarity is used. A non-local aggregation constraint term for smoothing image gradient graph is proposed, which is combined with the L0 norm minimization prior to obtain a new optimization framework. Then an alternating iteration algorithm is presented to efficiently solve the optimization problem of the new energy model. The non-local constraints of the new method weaken the ladder between similar blocks in a data-driven manner. Compared with the existing methods, the smoothing results of the new method are not only consistent, but also can keep the edges of the structure not shifting. 4. The application of image smoothing in intelligent image processing is studied. At the same time, it retains the structure information which is very important to human visual system and has a strong application value for content-related image processing. Firstly, the application of image smoothing in image zooming and zooming, image editing and other issues is studied. Then, a new multi-scale space construction method based on smoothing method is proposed and studied. The experimental results show that the image smoothing itself and the multi-scale space play a very good role in various applications.
【學(xué)位授予單位】：山東大學(xué)
【學(xué)位級別】：博士
【學(xué)位授予年份】：2016
【分類號】：TP391.41
，
本文編號：2235380
[Abstract]:The research goal of computer vision is to understand the scene through image or video. Image processing technology is the key technology to achieve this goal. Image denoising and smoothing are the basic problems in the field of image processing. Noise is unavoidably introduced into the image, resulting in image distortion. Image distortion will inevitably interfere with the follow-up work of image feature extraction and scene understanding, thus affecting the accuracy of computer intelligent processing tasks. Information, especially edges and texture details. If a good representation model can be established for a clean image, more useful information can be retained to ensure the accuracy of the restored image. Advanced computer vision tasks lay a solid foundation. Image denoising and image smoothing as the basis of image processing technology, in space projects, medicine, archaeology, industrial machine vision, military recognition, satellite image processing and other fields have a wide range of applications. There are a series of problems and challenges. Firstly, relying on the existing mathematical tools can not accurately describe the image, the existing methods based on various assumptions to establish the image representation model, there are certain limitations; secondly, the human visual system working mechanism is very complex, the current research on the principle of human perception is still in its infancy. It is impossible to define visually meaningful features in an image by a definite mathematical model. This paper focuses on the research hotspots and difficulties of image denoising and smoothing, and proposes an image processing method based on non-local sparseness. The main work includes: 1. An adaptive sparse coding denoising method based on PCA dictionary is proposed. By analyzing the statistical characteristics of sparse coding errors in PCA dictionary, Laplace function is used to approximate the distribution of coding errors and a posterior estimation theory is proposed. A new non-local sparse coding model is proposed in which the regularization parameters for balancing fidelity terms and non-local constraints are determined adaptively. To obtain reliable sparse coding estimation, an iterative shrinkage algorithm based on filtering is proposed. Filtering can effectively suppress the noise in the backward projection process and further obtain sparse coding. Robust estimation. The new method effectively improves the coding accuracy and achieves good texture preservation and noise removal. 2. An image smoothing method based on low rank sum gradient sparseness is proposed. Then a new image smoothing optimization method is proposed based on the global gradient sparse priori of smoothing image. The low rank priori constrains the strong correlation of each image block in the similar grouping of smoothing image. It can remove the small scale noise and details, maintain the edge of slender structure and ensure the consistent smoothing effect. A detailed flow chart based on alternating iteration approximation algorithm is given to solve the optimization problem. The new method can achieve high uniform smoothing effect and maintain important structural edges while removing details. 3. A non-local gradient clustering image smoothing method is proposed. The characteristics of natural image gradient map are analyzed and non-local self-similarity is used. A non-local aggregation constraint term for smoothing image gradient graph is proposed, which is combined with the L0 norm minimization prior to obtain a new optimization framework. Then an alternating iteration algorithm is presented to efficiently solve the optimization problem of the new energy model. The non-local constraints of the new method weaken the ladder between similar blocks in a data-driven manner. Compared with the existing methods, the smoothing results of the new method are not only consistent, but also can keep the edges of the structure not shifting. 4. The application of image smoothing in intelligent image processing is studied. At the same time, it retains the structure information which is very important to human visual system and has a strong application value for content-related image processing. Firstly, the application of image smoothing in image zooming and zooming, image editing and other issues is studied. Then, a new multi-scale space construction method based on smoothing method is proposed and studied. The experimental results show that the image smoothing itself and the multi-scale space play a very good role in various applications.
【學(xué)位授予單位】：山東大學(xué)
【學(xué)位級別】：博士
【學(xué)位授予年份】：2016
【分類號】：TP391.41
，
本文編號：2235380
本文鏈接：http://sikaile.net/shoufeilunwen/xxkjbs/2235380.html
最近更新
教材專著
