search

CA-USTC / License_Plate_Recognition_pytorch

该工作由电子科技大学陈昂、刘俊凯、夏子寒同学完成。我们提出可以使用深度学习的方案对中国车牌进行检测和识别。我们提出使用YOLOv3网络进行车牌检测，然后创新性地对检测后的车牌使用空间变换网络STN加以校正，最后使用LPRNet网络进行车牌的字符与数字识别。目前实测结果，在老师提供的45张数据集上，我们的YOLOv3网络检测准确率（IOU）达到98.2%，深度学习级联网络识别准确率为95.6%。我们采用大量测试集，最终我们的YOLOv3_STN_LPRNet级联网络识别准确率稳定在93.3%，不加空间变换网络STN的话，识别准确率在66%左右。 总体来说，使用深度学习的方案比传统方案效果提升的非常好，而且我们加入空间变换网络STN的做法对于提升识别准确率很有效。
72 stars 14 forks source link

NICE WORK !!! I referred to your STNet implementation and obtained a better license plate recognition algorithm #6

Open zjykzj opened 3 weeks ago

zjykzj commented 3 weeks ago

@CA-USTC Thank you very much for your open-source implementation. Through experiments and training, I have found that adding STNet can effectively improve the performance of license plate recognition.

Model ARCH Input Shape GFLOPs Model Size (MB) ChineseLicensePlate Accuracy (%) Training Data Testing Data
CRNN CONV+GRU (3, 48, 168) 4.0 58 82.147 269,621 149,002
CRNN_Tiny CONV+GRU (3, 48, 168) 0.3 4.0 76.590 269,621 149,002
LPRNetPlus CONV (3, 24, 94) 0.5 2.3 63.546 269,621 149,002
LPRNet CONV (3, 24, 94) 0.3 1.9 60.105 269,621 149,002
LPRNetPlus+STNet CONV (3, 24, 94) 0.5 2.5 72.130 269,621 149,002
LPRNet+STNet CONV (3, 24, 94) 0.3 2.2 72.261 269,621 149,002

The relevant code has been open sourced and can be viewed at zjykzj/crnn-ctc