發(fā)布時(shí)間：2018-02-19 22:05

  本文關(guān)鍵詞： 圖像去噪 核范數(shù)最小化 車(chē)牌傾斜 低秩紋理 仿射變換　出處：《合肥工業(yè)大學(xué)》2016年碩士論文　論文類(lèi)型：學(xué)位論文

【摘要】：近年來(lái)交通需求的日益增大,道路交通存在的矛盾越來(lái)越尖銳,交通的智能化是整個(gè)交通生態(tài)環(huán)境中必不缺少的一課。而其中應(yīng)用最廣的自動(dòng)車(chē)牌識(shí)別技術(shù)更是重要的一環(huán),為了給車(chē)牌識(shí)別提供更有質(zhì)量的圖像源,文章研究其中兩個(gè)方面的問(wèn)題,即車(chē)牌圖像的可視性增強(qiáng)和傾斜校正。首先在車(chē)牌圖像的可視性增強(qiáng)方面,本文主要針對(duì)圖像的噪聲以消除其所帶來(lái)的視覺(jué)影響。在圖像去噪非局部框架中,利用低秩模型的核范數(shù)最小化并將圖像的先驗(yàn)信息用權(quán)重系數(shù)的手段體現(xiàn)出來(lái)。文章中利用了矩陣較大奇異值在作用上具有高優(yōu)先級(jí)這一性質(zhì)并給較大奇異值賦予較小權(quán)值以達(dá)到較小的收縮,提出加權(quán)核范數(shù)最小化圖像去噪方法。通過(guò)實(shí)驗(yàn)論證,該方法在主觀視覺(jué)上整體和細(xì)節(jié)都獲得了較之前方法的優(yōu)勢(shì),而且其峰值信噪比也得到了較高的值。車(chē)牌識(shí)別中的另一方面,圖像傾斜同樣是字符分割乃至識(shí)別中重要的一步。基于此,提出一種適用范圍更廣的基于低秩紋理變換不變性的車(chē)牌傾斜校正方法。先將車(chē)牌圖像看成一個(gè)矩陣,然后利用對(duì)稱(chēng)規(guī)則結(jié)構(gòu)的低秩特性把經(jīng)過(guò)旋轉(zhuǎn)、仿射等變換的圖像恢復(fù)成低秩紋理。根據(jù)不同的環(huán)境條件,對(duì)隨機(jī)拍攝的傾斜車(chē)牌圖像進(jìn)行校正,實(shí)驗(yàn)表明即使面對(duì)一些如無(wú)邊框、車(chē)牌污染、環(huán)境光遮蔽、噪聲大等惡劣條件,文章的方法也得到了出色的校正結(jié)果。文章提出了一種借助于圖像整體信息進(jìn)行傾斜校正的新思路,由于該方法摒棄了以往的從角點(diǎn)、邊緣等局部特征的角度,因此具有更高的精確度和魯棒性。實(shí)驗(yàn)結(jié)果也證實(shí)了文章中方法的有效性且滿(mǎn)足整個(gè)識(shí)別系統(tǒng)的實(shí)時(shí)要求。
[Abstract]:In recent years, with the increasing demand for traffic, the contradiction of road traffic is becoming more and more acute. The intelligent traffic is a necessary lesson in the whole traffic ecological environment, and the most widely used automatic license plate recognition technology is an important part. In order to provide a better quality image source for license plate recognition, two problems are studied in this paper, namely, the visibility enhancement and tilt correction of license plate image. This paper mainly aims at the noise of image to eliminate the visual influence. In the non-local frame of image denoising, The kernel norm of the low rank model is minimized and the prior information of the image is represented by the weight coefficient. In this paper, the property that the larger singular value of the matrix has a high priority in action is used and the larger singular value is assigned to the larger singular value. To give a smaller weight value to achieve a smaller contraction, A weighted kernel norm minimization method for image denoising is proposed. The experimental results show that the proposed method has the advantages of both overall and detailed subjective vision. On the other hand, image tilt is also an important step in character segmentation and recognition. In this paper, a more widely used method for license plate tilt correction based on the invariance of low rank texture transform is proposed. Firstly, the license plate image is regarded as a matrix, and then the image is rotated by using the low rank characteristic of the symmetric regular structure. The image of affine transform is restored to low rank texture. According to different environmental conditions, the random image of inclined license plate is corrected. The experiment shows that even in the face of some things such as no border, license plate pollution, environmental light masking, In this paper, a new method for skew correction with the help of the whole information of the image is proposed, because the former corner point is abandoned by this method. Because of the angle of local features such as edges, it has higher accuracy and robustness. The experimental results also show that the proposed method is effective and meets the real-time requirements of the whole recognition system.
【學(xué)位授予單位】：合肥工業(yè)大學(xué)
【學(xué)位級(jí)別】：碩士
【學(xué)位授予年份】：2016
【分類(lèi)號(hào)】：TP391.41
，

本文編號(hào)：1518097