【摘要】 汽车手柄字符常采用人工检测方法进行缺陷检测，当人眼出现视觉疲劳时，会出现误检、漏检的现象。针对检测准确率低的问题，提出一种基于改进LeNet-5的深度学习字符缺陷检测算法。该算法首先对图像进行预处理和字符分割，然后调整分割后的字符大小并进行分类，制作成数据集。再调整LeNet-5网络的输入图像大小，增加输入的特征信息；引入批量归一化操作，提高网络的泛化能力；在反向传播过程中引入Adam优化器，提高参数更新的结果。最后使用改进网络训练模型进行实验。实验结果表明：在迭代2 000次的条件下，准确率为98.89%，和LeNet-5卷积神经网络相比准确率得到了提高。更多还原

【Abstract】 Automotive handle characters are often detected by manual inspection methods for defect detection. When the human eye experiences visual fatigue,false detection and missed detection will occur. To solve the problem of low detection accuracy,a deep learning character defect detection algorithm based on improved LeNet-5 is proposed. At first,image pre-processing and character segmentation are performed. Then the segmented characters are adjusted and classified to make a data-set. Then the input image size of LeNet-5 network is adjusted to increase the input feature information;the batch normalization operation is introduced to improve the generalization ability of the network;the Adam optimizer is introduced in the back-propagation process to improve the results of parameter updates. Finally,experiments are conducted by using the improved network training model. The experimental results show that the accuracy rate of this paper is 98.89%under the condition of 2 000 iterations. The accuracy is improved compared with LeNet-5 convolutional neural network.更多还原