發(fā)布時(shí)間：2018-12-13 07:03

【摘要】：數(shù)字圖像修復(fù)是利用圖像完好區(qū)域的信息對(duì)缺損區(qū)域進(jìn)行修復(fù)的一個(gè)過(guò)程。它要求修復(fù)后的結(jié)果保持良好的視覺(jué)觀賞性,最大限度地減少修復(fù)過(guò)程中的人為痕跡。數(shù)字圖像修復(fù)可以分為基于圖像自身內(nèi)容的修復(fù)技術(shù)和基于其它素材的多選擇圖像修復(fù)技術(shù)。相比于基于圖像自身內(nèi)容的修復(fù)技術(shù),基于其它素材的多選擇圖像修復(fù)技術(shù)有著明顯的優(yōu)勢(shì)。一方面,它可以有效地解決圖像大面積缺損問(wèn)題。另一方面,它為缺損圖像提供了更多的修復(fù)選擇,讓修復(fù)結(jié)果更加有多樣性和創(chuàng)造性。在媒體時(shí)代的今天,很多圖像處理技術(shù)在媒體應(yīng)用中發(fā)揮了重要的作用,基于其它素材的多選擇修復(fù)技術(shù)作為圖像處理中的重要技術(shù)為媒體應(yīng)用提供了重要的技術(shù)支持。它的基本流程可以歸納為以下三個(gè)步驟：首先,需要從海量的圖像庫(kù)中檢索出素材圖像(和待修復(fù)圖像相同或者相近的可以用來(lái)作為修復(fù)素材的圖像)；其次,需要對(duì)素材圖像中的素材區(qū)域(素材圖像中可以用來(lái)完成修復(fù)的目標(biāo)區(qū)域)進(jìn)行提��；再次,需要利用素材區(qū)域?qū)θ睋p圖像進(jìn)行修復(fù)。但是,如何從大數(shù)據(jù)圖像庫(kù)中選取大量且精準(zhǔn)的素材圖像,如何從圖像中精確地提取素材區(qū)域,如何減少修復(fù)過(guò)程中的人為痕跡(修復(fù)邊緣梯度變化明顯,前景和背景在顏色,紋理,噪聲等特征上有著明顯的差異)都是修復(fù)過(guò)程中的關(guān)鍵問(wèn)題。傳統(tǒng)的多選擇圖像修復(fù)算法往往沒(méi)有有效地解決這些問(wèn)題,從而直接降低了修復(fù)的質(zhì)量。本文針對(duì)以上這些問(wèn)題進(jìn)行了深度分析,對(duì)“基于檢索的多選擇圖像修復(fù)”進(jìn)行了深入研究,全文的主要貢獻(xiàn)包括以下幾點(diǎn)：(1)針對(duì)素材圖像檢索正確率低下的問(wèn)題,提出了一種素材圖像優(yōu)選模型。首先,該方法利用不同類(lèi)別間的聯(lián)合分布概率對(duì)素材圖像進(jìn)行初步檢索,利用文字信息提前刪除掉一些無(wú)關(guān)圖像,有效地減少了工作量。其次,利用K-means聚類(lèi)技術(shù)對(duì)圖像進(jìn)行分類(lèi),將圖像分為風(fēng)景類(lèi)圖像和物體類(lèi)圖像,根據(jù)圖像類(lèi)別利用不同的特征算子進(jìn)行計(jì)算。最后,利用改進(jìn)的空間金字塔函數(shù)完成待修復(fù)圖像和待檢索圖像的匹配,實(shí)現(xiàn)高精度的圖像匹配。三個(gè)環(huán)節(jié)緊密聯(lián)系,有效地解決了素材圖像檢索正確率低下的問(wèn)題,為多選擇的圖像修復(fù)技術(shù)提供了高效準(zhǔn)確的素材圖像。(2)針對(duì)難以精確提取素材區(qū)域的問(wèn)題,提出了一種優(yōu)化的素材區(qū)域提取模型。首先,該方法利用多尺度細(xì)節(jié)保留技術(shù)和多層平滑技術(shù)對(duì)素材圖像進(jìn)行深度優(yōu)化,使得素材區(qū)域和背景部分有著顯著的區(qū)分度。與此同時(shí),素材區(qū)域自身的顏色,紋理差異也明顯變小。優(yōu)化后圖像中的素材區(qū)域可以非常方便地被提取,有效地提升了提取精度。對(duì)于高質(zhì)量的摳像技術(shù)(特殊情況下,素材區(qū)域需要高精度的提取),本文利用訓(xùn)練樣本優(yōu)選模型進(jìn)行高精度摳取,有效地解決了摳取精度不足的問(wèn)題。新的模型有效地解決了無(wú)法精確提取素材區(qū)域的問(wèn)題,為多選擇的圖像修復(fù)技術(shù)提供了精準(zhǔn)的素材區(qū)域。(3)針對(duì)修復(fù)過(guò)程復(fù)雜,修復(fù)結(jié)果人為痕跡明顯的問(wèn)題,提出了一種高質(zhì)量的修復(fù)方法。首先,利用改進(jìn)的foe(field of experts)算法對(duì)待修復(fù)部分進(jìn)行修善,有效地提升了修復(fù)質(zhì)量。在修善過(guò)程中,該方法對(duì)訓(xùn)練圖像進(jìn)行優(yōu)選,有效地減少了工作量。然而,基于foe算法的簡(jiǎn)單修復(fù)缺乏多樣性和創(chuàng)意性,除此之外,我們利用前面技術(shù)檢索到的素材區(qū)域進(jìn)行多選擇性修復(fù)。在修復(fù)過(guò)程中,本文利用了多尺度的空間顏色匹配技術(shù),多尺度的空間紋理匹配技術(shù),以及多尺度的空間噪聲處理技術(shù)。這些方法可以使得素材區(qū)域和待修復(fù)圖像間的顏色,紋理等特征更加協(xié)調(diào),使修復(fù)后的圖像有更好的視覺(jué)效果。以上幾個(gè)環(huán)節(jié)緊密聯(lián)系,有效地優(yōu)化了修復(fù)結(jié)果。
[Abstract]:Digital image repair is a process of repairing the defect area with the information of the image intact area. It requires good visual appreciation of the results after the restoration, and the artificial marks in the repair process can be reduced to the maximum extent. Digital image repair can be divided into a repair technique based on the content of the image itself and a multi-selection image repair technique based on other materials. Compared with the image-based self-content-based repair technology, the multi-selection image restoration technology based on other materials has obvious advantages. in one aspect, that invention can effectively solve the problem of large-area defect of the image. On the other hand, it provides more repair options for the defective image, making the results more diverse and creative. In the media age, many image processing techniques play an important role in the media application, and the multi-selection and repair technology based on other materials provides important technical support for media application as an important technique in image processing. its basic flow can be summed up as three steps: first, it is necessary to retrieve the material image (the same or similar image as the image to be repaired) from the mass of the image library; secondly, it is necessary to extract the material area in the material image (the target area that can be used to complete the repair in the material image); once again, the defect image needs to be repaired with the material area. However, how to select a large number of and accurate material images from the large data image library, how to extract the material area accurately from the image, how to reduce the artifacts in the repair process (the change of the repair edge gradient is clear, the foreground and the background are in color, texture, There is a significant difference in the characteristics of noise, etc.), which is the key problem in the repair process. Traditional multi-selection image restoration algorithms often do not effectively solve these problems, thus directly reducing the quality of the repair. In this paper, the depth analysis of the above problems is carried out, and the 鈥淢ulti-selection image restoration based on retrieval鈥，

本文編號(hào)：2376121