【摘要】：帶鋼產(chǎn)品,作為鋼鐵產(chǎn)品中的一個重要部分,已經(jīng)成為航空航天、機械制造、汽車生產(chǎn)、化工等工業(yè)的重要原材料,其質(zhì)量直接影響著產(chǎn)品的最終性能。因此,作為帶鋼表面質(zhì)量自動評估的一種重要手段,基于機器視覺的帶鋼表面缺陷在線檢測的研究具有重要的理論與現(xiàn)實意義。針對現(xiàn)有常規(guī)缺陷檢測方法在實際應用中吞吐量低、檢測準確率不高的問題,本文從基于紋理異常檢測和機器學習兩方面著手,分別提出了基于局部二值模式(LBP)的缺陷紋理異常檢測和基于卷積神經(jīng)網(wǎng)絡(luò)CNN的缺陷檢測方法。同時,為了提高分類算法的準確性,提出一種基于特征對的改進ReliefF特征選擇方法。本文研究的內(nèi)容與成果如下:(1)針對傳統(tǒng)檢測方法在帶鋼缺陷檢測中誤檢率和漏檢率較高,檢測算法參數(shù)較多的問題,提出了一種基于多尺度LBP編碼的缺陷檢測方法。在不同尺度下對帶鋼表面圖像創(chuàng)建高斯差分金字塔模型,確定缺陷疑似區(qū)域,然后針對可疑區(qū)域進行經(jīng)過閾值化處理后的LBP編碼,最后將所有尺度下的編碼圖像融合,生成最終的像素級缺陷位置信息,將檢測結(jié)果中的連通域合并(ROI合并),生成完整的缺陷位置信息。(2)為了排除背景紋理的干擾,抑制偽缺陷的產(chǎn)生,本文再次提出一種基于奇異值分解的缺陷檢測方法SVD-LBPH,通過對待檢測圖像進行SVD分解與重構(gòu),弱化背景紋理,然后采用LBP對圖像進行編碼,提取LBP直方圖相關(guān)的統(tǒng)計特征,通過將特征與設(shè)定的閾值比較,最終檢測出缺陷。(3)實時檢測對實時性要求較高,實時檢測結(jié)束后,需要在檢測出來的缺陷區(qū)域進行即時檢測,進一步剔除偽缺陷。在即時檢測階段,針對深度學習算法在目標檢測與圖像分類的高準確率,采用卷積神經(jīng)網(wǎng)絡(luò)CNN進行缺陷的檢測,通過與其它基于機器學習的缺陷檢測方法進行對比,驗證了卷積神經(jīng)網(wǎng)絡(luò)在缺陷檢測的高效潛力。(4)為了對缺陷類別進行辨別,本文提取了缺陷的灰度特征、灰度共生陣以及頻域等特征。為了提高分類準確率,剔除不相關(guān)的特征,同時為了避免因特征維數(shù)過大而造成的過擬合,采用了特征篩選的手段。本文利用更新特征對權(quán)重的方式對ReliefF進行改進,實現(xiàn)特征的降維。最后利用SVM對篩選后的特征進行分類。
[Abstract]:Strip products, as an important part of iron and steel products, have become an important raw material in aerospace, mechanical manufacturing, automobile production, chemical industry, etc. The quality of strip products directly affects the final performance of the products. Therefore, as an important means of automatic evaluation of strip surface quality, the research of on-line detection of strip surface defects based on machine vision has important theoretical and practical significance. Aiming at the problems of low throughput and low detection accuracy of the existing conventional defect detection methods in practical applications, this paper starts from two aspects: texture-based anomaly detection and machine learning. Defect texture anomaly detection based on local binary mode (LBP) and defect detection based on convolution neural network CNN are proposed respectively. At the same time, in order to improve the accuracy of the classification algorithm, an improved ReliefF feature selection method based on feature pairs is proposed. The contents and achievements of this paper are as follows: (1) aiming at the problems of high false detection rate and high miss detection rate and more parameters of detection algorithm, a new defect detection method based on multi-scale LBP coding is proposed. Gao Si differential pyramid model is created for the strip surface image at different scales to determine the suspected defect area, and then the LBP coding after threshold processing is carried out for the suspected area. Finally, the coding image at all scales is fused. The final pixel level defect location information is generated, and the connected domain (ROI merging) in the detection result is combined to generate the complete defect location information. (2) in order to eliminate the interference of background texture, the false defect is suppressed. In this paper, a new defect detection method based on singular value decomposition (SVD) is proposed, which weakens background texture by SVD decomposition and reconstruction of detected image, and then uses LBP to encode the image and extract the statistical features related to LBP histogram. By comparing the features with the set threshold, the defects are finally detected. (3) Real-time detection requires high real-time performance. After the real-time detection, it is necessary to detect the defects in the detected areas immediately, and further eliminate the false defects. In the phase of immediate detection, aiming at the high accuracy of depth learning algorithm in target detection and image classification, a convolutional neural network (CNN) is used to detect defects, which is compared with other defect detection methods based on machine learning. The high efficiency potential of convolution neural network in defect detection is verified. (4) in order to distinguish the defect category, the gray level feature, gray level co-occurrence matrix and frequency domain feature of defect are extracted in this paper. In order to improve classification accuracy, eliminate irrelevant features, and avoid over-fitting caused by large feature dimension, feature selection is adopted. In this paper, ReliefF is improved by updating the weight of feature pair to reduce the dimension of feature. Finally, SVM was used to classify the selected features.
【學位授予單位】：河北工業(yè)大學
【學位級別】：碩士
【學位授予年份】：2015
【分類號】：TG142.15;TP391.41

【參考文獻】

相關(guān)期刊論文 前10條

1 徐科;宋暢;;基于全局二值模式的特征提取方法及其應用[J];模式識別與人工智能;2013年09期

2 陳躍;張曉光;;帶鋼表面缺陷圖像的可拓分類算法[J];計算機工程與應用;2013年21期

3 宋克臣;顏云輝;彭怡書;董德威;;引入局部信息的帶鋼缺陷圖像凸優(yōu)化活動輪廓分割模型[J];機械工程學報;2012年20期

4 許豪;孔建益;湯勃;王興東;劉源l，

本文編號：2385177