Abstract: Color difference detection of polycrystalline wafers with complex textures was a challenge in the manufacturing processes of solar cells. The traditional algorithm in color difference detection was not suitable for situations where the categories of color differences varied greatly, and the classification results were inaccurate. A detection algorithm of multi-component convolution neural network was proposed based on color features of different components. It is found that color features are different in H, S and V components by analyzing the feature distribution of polysilicon wafer images in HSV color spaces. The influences of model depth and convolution kernel size on detection results were evaluated to build the best convolution neural network structure based on full convolution neural network. A multi-component convolution neural network model was constructed based on the best network model to enhance the ability to distinguish different color difference features. The experimental results show that the accuracy, the values of MCC and F1Score of the multi-component convolution neural network are 92.28%, 95.45%, and 94.03% respectively, which has higher detection accuracy than that of other algorithms.

Key words: polycrystalline wafer, color difference, multi-component, convolutional neural network

摘要： 具有复杂纹理的多晶硅晶片颜色差异检测是太阳能电池片制造过程中的一个挑战。针对传统的色差检测算法不适用于颜色差异类别变化大的场合，且分类结果不精确的问题，基于不同分量的颜色特征提出了一种多分量卷积神经网络的检测算法。通过分析多晶硅晶片图像在HSV颜色空间的特征分布，发现颜色特征在H、S和V分量中表现不同；基于全卷积神经网络，通过评估模型深度和卷积核尺寸大小对检测结果的影响来搭建最佳的网络结构；为了增强对不同颜色差异特征的区分能力，基于最佳的网络模型，构建了一个多分量的卷积神经网络模型。实验结果表明，多分量卷积神经网络的准确率、MCC值和F1Score值分别为92.28%、95.45%和94.03%，相比其他算法具有更高的检测精度。

关键词: 多晶硅晶片, 颜色差异, 多分量, 卷积神经网络