(1. 燕山大學 國家冷軋板帶裝備及工藝工程技術(shù)研究中心，秦皇島 066004；
2. 燕山大學 亞穩(wěn)材料制備技術(shù)與科學國家重點實驗室，秦皇島 066004)
摘 要: 表面質(zhì)量是冷軋銅帶重要質(zhì)量指標之一。為實現(xiàn)銅帶表面缺陷的精準自動檢測，首先對常見表面缺陷進行分類，并制作了銅帶表面缺陷圖像數(shù)據(jù)集(YSU_CSC)；然后，以卷積神經(jīng)網(wǎng)絡(luò)EfficientNet為核心，基于遷移學習策略，通過訓練實驗建立了冷軋銅帶表面缺陷智能識別模型，同時與其他三種常用的卷積神經(jīng)網(wǎng)絡(luò)缺陷識別結(jié)果進行對比。結(jié)果表明：該模型的精度較高，準確率達到93.05%，單張缺陷圖像平均識別時間為197 ms，綜合性能較好，可以滿足工程要求；最后，將該模型在測試集上的缺陷識別結(jié)果進行可視化，分析了該模型對部分圖像識別錯誤的原因，給出了進一步優(yōu)化的方向。
關(guān)鍵字: 冷軋銅帶；表面缺陷；卷積神經(jīng)網(wǎng)絡(luò)；遷移學習；識別模型
（1. National Engineering Research Center for Equipment and Technology of Cold Rolling Strip, Yanshan University, Qinhuangdao 066004, China;
2. State Key Laboratory of Metastable Materials Science and Technology, Yanshan University, Qinhuangdao 066004, China）
Abstract:Surface quality is one of the significant indicators of cold rolling copper strip product quality. In order to realize the accurate and automatic detection of copper strip surface defects, this article first classifies common surface defects, and creates a copper strip surface defect image dataset (YSU_CSC). Then, with the EfficientNet convolutional neural network as the core, based on the transfer learning strategy, the optimal cold rolling copper strip surface defect recognition model was established through training experiments. At the same time, it was compared with the defect recognition model established by the other three convolution neural network algorithms. The results show that the accuracy of this model is the highest, and the accuracy reaches 93.05%, the average recognition time of single defect image is 197 ms, and the comprehensive performance is the best, which basically meets the engineering requirements. Finally, the defect recognition results of this model on the testing set were visualized, the causes of the error of defect recognition was analyzed, and the direction of further optimization is given.
Key words: cold rolling copper strip; surface defects; EfficientNet CNN; transfer learning; recognition model
