史進(jìn)偉， 郭朝勇， 劉紅寧
（軍械工程學(xué)院基礎(chǔ)部，河北 石家莊 050003）
摘  要：為實(shí)現(xiàn)槍彈外觀缺陷自動(dòng)檢測(cè)，提出一種基于BP神經(jīng)網(wǎng)絡(luò)的槍彈外觀缺陷自動(dòng)識(shí)別與分類方法。首先針對(duì)槍彈外觀缺陷圖像特點(diǎn)，從形狀、顏色、紋理提取類別差異明顯的缺陷特征向量，作為神經(jīng)網(wǎng)絡(luò)的輸入，以提高分類效果；然后通過(guò)經(jīng)驗(yàn)和實(shí)驗(yàn)驗(yàn)證確定神經(jīng)網(wǎng)絡(luò)結(jié)構(gòu)及參數(shù)，并分析傳統(tǒng)BP算法在槍彈外觀缺陷分類應(yīng)用中的不足，通過(guò)優(yōu)化BP算法以提高網(wǎng)絡(luò)分類性能。實(shí)驗(yàn)表明：優(yōu)化BP算法能夠有效分類槍彈外觀缺陷測(cè)試樣本，識(shí)別率達(dá)到92.1%，與傳統(tǒng)BP算法相比，提高了收斂速度，并表現(xiàn)出較好的準(zhǔn)確性和魯棒性，能夠更好滿足槍彈外觀缺陷自動(dòng)檢測(cè)要求。
關(guān)鍵詞：槍彈外觀缺陷；特征提取；BP神經(jīng)網(wǎng)絡(luò)；識(shí)別與分類
中圖分類號(hào)：TP391.4；TJ411；TJ06；TP274+.2        文獻(xiàn)標(biāo)志碼：A       文章編號(hào)：1674-5124（2013）04-0026-05
Identification and classification of bullet surface defect based on BP neural network
SHI Jin-wei， GUO Chao-yong， LIU Hong-ning
（Department of Basic Courses，Ordnance Engineering College，Shijiazhuang 050003，China）
Abstract: In order to achieve automatic detection of bullet surface defect， a new method is proposed for automatically identifying and classifying the bullet surface defects on the basis of BP neural network. Firstly， according to the property of bullet surface defects， the distinct defect feature vector is extracted as the import of neural network from shape， color and texture. Secondly， the structure and parameter of neural network are ascertained by experience and experiment confirmation， the disadvantage of bullet surface defect classification by BP neural network is analyzed， and the classification capability of network is improved by optimized BP method. The experimental results show that the test stylebook of bullet surface defect can be classified by the optimized BP method effectively， and the discriminating rate can reach 92.1%. In the experiment of contrast with traditional BP method， the speed of convergence is improved， and the new method has a good ability of accuracy and robustness， can better satisfy the need of automatic detection of bullet surface defects.
Key words: bullet surface defect； feature extraction； BP neural network； identification and classification