基于堆栈式自编码网络的电子线路分类算法

Electronic circuit classification algorithm based on stacked auto-encoders network

由于电子线路细节特征种类多且与背景细节特征相似度较高,基于图像颜色、纹理形状等低层特征的分类算法不能满足高精度分类的需求。针对是否具有数字背景的电子线路分类问题,利用深度学习方法堆栈式降噪自编码网络以及方向梯度直方图特征提取算法对1 840张工业电子线路图片的分类进行研究。实验结果表明,对缩放到68×68大小的电子线路图像进行去均值、归一化及白化等预处理能有效降低不同光照强度的影响,同时降低了像素间的相关性,因此在后续训练过程中能得到更加具有分类代表性的特征使分类的准确率提高约6%;预处理后提取图片1 152维的方向梯度直方图特征作为输入,通过两层隐含层降噪自编码训练及反向传播权值微调后能更加准确、稳定地区分出具有数字背景的电子线路。

Because of the difficulty to accurately distinguish the minutiae of the electronic circuit and the background details based on the low level features including color and texture,and aiming at the classification problem of the electronic circuit with digital background,this paper used the deep learning method of stacked denoising auto-encoders network(SDAE)and histogram of oriented gradients(HOG)algorithm to study the classification result of 1 840 industrial electronic circuit images.The experiment results indicate that,using the electronic circuit images pretreated by meaning,normalization and whitening can obviously reduce the interferences of light intensity and correlation between pixels,therefore,more representative features can be obtained in the subsequent training process;after pretreatment,extracting 1 152 dimensional HOG feature vectors as the inputs of the SDAE network,and training the two layers of the SDAE network with back propagation method can more accurately and steadily classify the electronic circuit with digital background.