基于BP神經(jīng)網(wǎng)絡(luò)的玻璃缺陷識(shí)別技術(shù)研究

摘 要:為了準(zhǔn)確、快速地對(duì)玻璃質(zhì)量進(jìn)行分類，提出一種基于BP神經(jīng)網(wǎng)絡(luò)的玻璃缺陷識(shí)別方法。由于不變矩與灰度共生矩陣分別可以描述圖像的形狀與紋理，在分析了缺陷灰度圖像特點(diǎn)的基礎(chǔ)上，將圖像的紋理特征和不變矩特征融合，綜合提取出一個(gè)分類能力更強(qiáng)的特征向量，再使用一個(gè)擬牛頓改進(jìn)算法的三層前向BP網(wǎng)絡(luò)。作為分類器，對(duì)常見(jiàn)的玻璃缺陷進(jìn)行了識(shí)別。通過(guò)實(shí)驗(yàn)對(duì)比該方法和傳統(tǒng)的單一特征識(shí)別法，證明該方法不僅具有更高的識(shí)別率，并且實(shí)時(shí)性較好，為玻璃缺陷的自動(dòng)識(shí)別提供了一種新的途徑。

關(guān)鍵詞:玻璃缺陷; BP神經(jīng)網(wǎng)絡(luò); 不變矩; 灰度共生矩陣

中圖分類號(hào):TP919.8 文獻(xiàn)標(biāo)識(shí)碼:A

文章編號(hào):1004-373X(2010)14-0045-04

Research of Recognition Technology for Glass Defect Based on BP Neural Network

WANG Su-zhe， WANG Zhao-ba， JIN Yong， CHEN You-xing

(National Key Laboratory for Electronic Measurement Technology， North University of China， Taiyuan 030051， China)

Abstract: A method of recognizing glass defects based on BP neural network is put forward to classify the glass accurately and quickly. Because the invariant moment and gray-level co-occurrence matrix can describe the shape and texture of images respectively， a more classifiable characteristic vector is extracted comprehensively by fusing the two features of images on the basis of characteristic analysis of the images with defects， then a three-layer feed-forward propagation neural network based on BFGS algorithm is adopted as a classifier to realize the recognition and classification of the images with familiar glass defect. The experiment， in which this method was compared with the conventional way of single character recognition， demonstrates the superiority of higher recognition rate and better real-time performance， which offers a new approach to automatic recognition of glass defects.

Keywords: glass defect; BP neural network; invariant moment; gray-level co-occurrence matrix

0 引言

浮法玻璃生產(chǎn)工藝是世界上最先進(jìn)的玻璃制造工藝，但是在生產(chǎn)過(guò)程中，還是會(huì)不可避免地產(chǎn)生一些缺陷。常見(jiàn)的玻璃缺陷有以下幾種類型:氣泡、結(jié)石、劃痕、夾雜。這些缺陷嚴(yán)重影響了玻璃的外觀質(zhì)量，降低了玻璃的透光性能和機(jī)械強(qiáng)度。為了提高浮法玻璃質(zhì)量并方便劃分玻璃質(zhì)量的等級(jí)，必須對(duì)這些缺陷進(jìn)行分類。

目前，國(guó)外從20世紀(jì)80年代就開(kāi)始研究計(jì)算機(jī)視覺(jué)檢測(cè)技術(shù)，雖然至今已自主研制出一些玻璃自動(dòng)檢測(cè)設(shè)備，但是價(jià)格昂貴，且技術(shù)資料保密[1-2]。國(guó)內(nèi)也有類似設(shè)備和技術(shù)的研究，與國(guó)外相比，在檢測(cè)及識(shí)別速度、精度方面還存在著較大的差距[3-4]。因此，本文深入研究了玻璃檢測(cè)過(guò)程中的缺陷識(shí)別技術(shù)，旨在提高玻璃檢測(cè)系統(tǒng)的整體檢測(cè)水平。……