duanshengliu
commented
7 months ago First of all, I agree with your points of view. As you mentioned in the first point, I didn't consider the extreme lighting conditions, and the dataset mainly comes from the CCPD dataset, which contains many photos taken at night. However, I believe that this is a good idea to improve robustness. Additionally, I also believe that there are many advanced segmentation networks that will bring further improvement. I would be very happy to receive your pull request.
Dear Maintainers,
Firstly, I would like to extend my compliments on the remarkable work being carried out in the realm of Chinese plate recognition with the End-to-end-for-chinese-plate-recognition repository. The integration of U-Net with CNN for the purpose of plate detection and character recognition is indeed commendable.
Upon perusal of the project and its associated documentation, I have observed that the system performs admirably under a range of challenging conditions, such as skewed angles and exposure variances. However, I am curious about the potential avenues for further enhancing the robustness of the recognition system, particularly in scenarios of extreme lighting conditions, such as overexposure and underexposure that extend beyond typical environmental variations.
To this end, I propose the following points for discussion and potential development:
Adaptive Thresholding: The current methodology utilises a binary thresholding technique post-U-Net segmentation. Have we considered the implementation of adaptive thresholding mechanisms that could dynamically adjust to the lighting conditions of the input image?
Illumination Normalisation: Pre-processing steps that normalise the illumination of the input images could be beneficial. Techniques such as Histogram Equalisation or the application of the Retinex theory might prove to be efficacious in this regard.
Data Augmentation: To improve the model's exposure to various lighting conditions, could we potentially expand the training dataset with synthetically altered images that simulate these extreme conditions?
Model Enhancement: Are there any plans to explore advanced neural network architectures that might be inherently more resilient to such environmental factors?
I believe that addressing these points could significantly bolster the system's performance and its applicability in real-world scenarios. I would be keen to contribute to this discussion and assist in any subsequent development efforts.
I look forward to your thoughts on the matter.
Best regards, yihong1120