Validation images cannot be read properly even with high accuracy

[IAmGreyBunny]

Hi, i am trying to train an ocr model for recognising license plate. I am currently training with None-VGG-BiLSTM-CTC to be used with easyocr. i have a training dataset of 800 and a validation dataset of 200 generated by a script which creates varying background and font color. During training, the validation accuracy starts to go up at around 600 iterations and steadily increase to about 80+% after 1000 iterations however when i used it in the demo.py or in easyocr on some of the images from the same validation dataset that is being used during training, it gives random result that does not even come close to the actual word. I am new to AI and machine learning in general, any help would be appreciated.

Things ive tried: 1) changing valInterval(increasing and decreasing) 2) batch-size(from 192 to 32) 3) imgW and imgH

Below is my opt.txt for the training

------------ Options ------------- exp_name: None-VGG-BiLSTM-CTC-Seed1111 train_data: result/train valid_data: result/test manualSeed: 1111 workers: 4 batch_size: 32 num_iter: 300000 valInterval: 50 saved_model: FT: False adam: False lr: 1 beta1: 0.9 rho: 0.95 eps: 1e-08 grad_clip: 5 baiduCTC: False select_data: ['/'] batch_ratio: ['1'] total_data_usage_ratio: 1.0 batch_maxlength: 34 imgH: 64 imgW: 600 rgb: False character: 0123456789!"#$%&'()*+,-./:;<=>?@[]^`{|}~ €ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz sensitive: False PAD: False data_filtering_off: False Transformation: None FeatureExtraction: VGG SequenceModeling: BiLSTM Prediction: CTC num_fiducial: 20 input_channel: 1 output_channel: 256 hidden_size: 256 num_gpu: 1 num_class: 97

[ozlem-atiz]

hello .İ am try to training the licance plate . But I have a problem like you. did you manage

[IAmGreyBunny]

hello .İ am try to training the licance plate . But I have a problem like you. did you manage

Hi, unfortunately ive moved on from the project, never found a solution