Not much to say, just look at the result picture:
It can be seen that the recognition effect is still very good.
Half a year ago, I wrote [python + opencv + pytorch] license plate extraction, segmentation, and recognition to extract, segment, and recognize license plates. The extraction and segmentation of license plates were realized by OpenCV. However, my personal grasp of OpenCV is limited. Finally The effect of extracting the license plate is not very good, and the robustness is very poor. At that time, I thought that if other deep learning can be used to extract the ticket, the effect may be better. After querying a lot of information, maybe use the target detection algorithm YOLO in deep learning V3 is used to extract the license plate, but something suddenly happened in the middle of learning the YOLO V3 algorithm, and this matter was put on hold. Until the last few days, I suddenly wanted to implement this function on a whim, and then found that it might be more appropriate to use semantic segmentation to do this function. Because the results detected by YOLO V3 may contain a lot of noise, which can easily lead to the final result of license plate character recognition. If semantic segmentation is used to do it, the noise problem should be easier to solve, so I went to learn the semantic segmentation framework deeplab v3+, and then made a data set of 274 license plates (the data set is made with labelme, which is relatively simple, so I won’t go into details here)
Of course, the deeplab v3+ framework was not written by me. I used the one written by Bubbliiiiing . For details on how to implement it, please refer to the article of the big guy. Here I just used the data set I collected to run for a while. Later, on the basis of this license plate extraction, I will realize the license plate segmentation and finally realize the license plate recognition. The code project and data will not be released for the time being, and it should be released after all license plate recognition is realized in the future.