基于字符分割和LeNet-5网络的字符验证码识别

Character Verification Code Recognition Based on Character Segmentation and LeNet-5 Network

摘要:

为了解决传统验证码识别方法效率低，精度差的问题，设计了一种先分割后识别的验证码处理方案。该方案在预处理阶段用中值滤波去噪，再利用霍夫变换对图像字符进行矫正；在字符分割阶段，利用垂直投影算法确定验证码字符块个数，以及字符坐标点，再用颜色填充算法对验证码进行初步分割，根据分割后的字符块数量对粘连字符进行二次分割；在识别阶段，我们对LeNet-5网络进行了改进，修改了输入层，并用全连接层替换了LeNet-5网络中的C5层，以此来对验证码字符进行识别；实验表明，对于非粘连验证码和粘连验证码，单张图片分割时间为0.14和0.15ms，分割准确率为98.75%和97.25%，识别准确率为99.99%和97.7%；结果表明，该算法对验证码分割和识别都有着很好的效果。

Abstract:

To address the low efficiency and accuracy of traditional captcha recognition methods, we designed a captcha processing solution that involves segmentation and recognition stages. In the preprocessing stage, we applied median filtering for noise reduction and used the Hough transform to correct the image characters. In the character segmentation stage, we used the vertical projection algorithm to determine the number of character blocks and their coordinates, and then used the color filling algorithm for preliminary segmentation. We also performed a second segmentation for connected characters based on the number of segmented character blocks. In the recognition stage, we improved the LeNet-5 network by modifying the input layer and replacing the C5 layer with a fully connected layer for character recognition. Experimental results showed that for non-connected and connected captchas, the segmentation time for a single image was 0.14ms and 0.15ms, respectively, with segmentation accuracies of 98.75% and 97.25% and recognition accuracies of 99.99% and 97.7%. These results demonstrate that our algorithm has good performance for captcha segmentation and recognition.