CAJ | 학술논문

目前对太阳能电池片的缺陷检测仍依赖人工完成，很难通过传统的 CCD 成像系统自动识别。作为一种多层神经网络学习算法，深度学习因对输入样本数据强大的特征提取能力而受到广泛关注。文中提出一种基于深度学习的太阳能电池片表面缺陷检测方法，该方法首先根据样本特征建立深度置信网络(DBN)，并训练获取网络的初始权值；然后通过 BP 算法微调网络参数，取得训练样本到无缺陷模板之间的映射关系；最后利用重构图像与缺陷图像之间的对比关系，实现测试样本的缺陷检测。实验表明 DBN 能较好地建立上述映射关系，且准确、快速地进行缺陷检测。
목전대태양능전지편적결함검측잉의뢰인공완성，흔난통과전통적 CCD 성상계통자동식별。작위일충다층신경망락학습산법，심도학습인대수입양본수거강대적특정제취능력이수도엄범관주。문중제출일충기우심도학습적태양능전지편표면결함검측방법，해방법수선근거양본특정건립심도치신망락(DBN)，병훈련획취망락적초시권치；연후통과 BP 산법미조망락삼수，취득훈련양본도무결함모판지간적영사관계；최후이용중구도상여결함도상지간적대비관계，실현측시양본적결함검측。실험표명 DBN 능교호지건립상술영사관계，차준학、쾌속지진행결함검측。
Defects of solar cells are detected mainly by manual operation, and they are difficult to be detected automatically by traditional charge-coupled device ( CCD) imaging system. As a training multi-layer neural network, deep learning draws great attention due to its strong ability to extract features from input sample data. A method for solar cells surface defects detection based on deep learning is proposed. Firstly, deep belief networks(DBN) are established and trained according to the sample features to obtain the initial weights of the networks. Then, the traditional BP algorithm is conducted to fine-tune the network parameters to get the mapping relationship between the training samples and the defect-free template. Finally, the defects of testing samples are detected by the contrast between the reconstruction image and the defect image. Experimental results show that DBN perfectly establishes the mapping relationship, and it can quickly detect defects with a high accuracy.