paddle ocr 文本检测模型训练
- 2024-12-11 19:25:00
- pjd
- 原创 25
文本检测
我们基于PP-OCRv3使用3种方案 进行检测模型的训练、评估：
PP-OCRv3中文超轻量检测预训练模型
PP-OCRv3中文超轻量检测预训练模型+验证集padding
PP-OCRv3中文超轻量检测预训练模型+ finetune
#模型下载下载地址：https://aistudio.baidu.com/modelsdetail/17/download
1. 中文文本检测模型
模型名称模型简介推理模型大小下载地址
ch_PP-OCRv3_det原始超轻量模型，支持中英文、多语种文本检测3.80M 训练模型
2. 文本识别模型
模型名称模型简介推理模型大小下载地址
ch_PP-OCRv3_rec原始超轻量模型，支持中英文、数字识别12.4M 训练模型
说明： 训练模型是基于预训练模型在真实数据与竖排合成文本数据上finetune得到的模型，在真实应用场景中有着更好的表现，预训练模型则是直接基于全量真实数据与合成数据训练得到，更适合用于在自己的数据集上finetune。
3. 文本方向分类模型
模型名称模型简介推理模型大小下载地址
ch_ppocr_mobile_v2.0_cls
原始分类器模型，对检测到的文本行文字角度分类
1.38M
推理模型 / 训练模型
采用方案3：预训练模型 + finetune
在1500图片上进行训练和评估，其中train数据1200张，val数据300张，
修改配置文件configs/det/ch_PP-OCRv3/ch_PP-OCRv3_det_student.yml中的以下字段：
Global.epoch_num: 这里设置为1，方便快速跑通，实际中根据数据量调整该值
Global.save_model_dir：模型保存路径
Global.pretrained_model：指向预训练模型路径，'./pretrain_models/en_PP-OCRv3_det_distill_train/student.pdparams'
Optimizer.lr.learning_rate：调整学习率，本实验设置为0.0005
Train.dataset.data_dir：指向训练集图片存放目录,'/home/aistudio/dataset'
Train.dataset.label_file_list：指向训练集标注文件,'/home/aistudio/dataset/det_gt_train.txt'
Train.dataset.transforms.EastRandomCropData.size：训练尺寸改为[480,64]
Eval.dataset.data_dir：指向验证集图片存放目录,'/home/aistudio/dataset/'
Eval.dataset.label_file_list：指向验证集标注文件,'/home/aistudio/dataset/det_gt_val.txt'
Eval.dataset.transforms.DetResizeForTest：评估尺寸，添加如下参数
limit_side_len: 64
limit_type:'min'
#确认安装paddle
conda install paddlepaddle-gpu==2.6.2 cudatoolkit=11.6 -c https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/Paddle/ -c conda-forge
执行下面命令启动训练：
cd ~/PaddleOCR/
python tools/train.py -c configs/det/ch_PP-OCRv3/ch_PP-OCRv3_det_student.yml
(paddle_ocr) PS ~\PaddleOCR> python tools/train.py -c configs/det/ch_PP-OCRv3/ch_PP-OCRv3_det_student_cimc.yml [2024/12/11 13:35:07] ppocr INFO: Architecture : [2024/12/11 13:35:07] ppocr INFO: Backbone : [2024/12/11 13:35:07] ppocr INFO: disable_se : True [2024/12/11 13:35:07] ppocr INFO: model_name : large [2024/12/11 13:35:07] ppocr INFO: name : MobileNetV3 [2024/12/11 13:35:07] ppocr INFO: scale : 0.5 [2024/12/11 13:35:07] ppocr INFO: Head : [2024/12/11 13:35:07] ppocr INFO: k : 50 [2024/12/11 13:35:07] ppocr INFO: name : DBHead [2024/12/11 13:35:07] ppocr INFO: Neck : [2024/12/11 13:35:07] ppocr INFO: name : RSEFPN [2024/12/11 13:35:07] ppocr INFO: out_channels : 96 [2024/12/11 13:35:07] ppocr INFO: shortcut : True [2024/12/11 13:35:07] ppocr INFO: Transform : None [2024/12/11 13:35:07] ppocr INFO: algorithm : DB [2024/12/11 13:35:07] ppocr INFO: model_type : det [2024/12/11 13:35:07] ppocr INFO: Eval : [2024/12/11 13:35:07] ppocr INFO: dataset : [2024/12/11 13:35:07] ppocr INFO: data_dir : ./work/train_data [2024/12/11 13:35:07] ppocr INFO: label_file_list : ['./work/train_data/img/Label.txt'] [2024/12/11 13:35:07] ppocr INFO: name : SimpleDataSet [2024/12/11 13:35:07] ppocr INFO: transforms : [2024/12/11 13:35:07] ppocr INFO: DecodeImage : [2024/12/11 13:35:07] ppocr INFO: channel_first : False [2024/12/11 13:35:07] ppocr INFO: img_mode : BGR [2024/12/11 13:35:07] ppocr INFO: DetLabelEncode : None [2024/12/11 13:35:07] ppocr INFO: DetResizeForTest : None [2024/12/11 13:35:07] ppocr INFO: NormalizeImage : [2024/12/11 13:35:07] ppocr INFO: mean : [0.485, 0.456, 0.406] [2024/12/11 13:35:07] ppocr INFO: order : hwc [2024/12/11 13:35:07] ppocr INFO: scale : 1./255. [2024/12/11 13:35:07] ppocr INFO: std : [0.229, 0.224, 0.225] [2024/12/11 13:35:07] ppocr INFO: ToCHWImage : None [2024/12/11 13:35:07] ppocr INFO: KeepKeys : [2024/12/11 13:35:07] ppocr INFO: keep_keys : ['image', 'shape', 'polys', 'ignore_tags'] [2024/12/11 13:35:07] ppocr INFO: loader : [2024/12/11 13:35:07] ppocr INFO: batch_size_per_card : 1 [2024/12/11 13:35:07] ppocr INFO: drop_last : False [2024/12/11 13:35:07] ppocr INFO: num_workers : 2 [2024/12/11 13:35:07] ppocr INFO: shuffle : False [2024/12/11 13:35:07] ppocr INFO: Global : [2024/12/11 13:35:07] ppocr INFO: cal_metric_during_train : False [2024/12/11 13:35:07] ppocr INFO: checkpoints : None [2024/12/11 13:35:07] ppocr INFO: debug : True [2024/12/11 13:35:07] ppocr INFO: distributed : False [2024/12/11 13:35:07] ppocr INFO: epoch_num : 500 [2024/12/11 13:35:07] ppocr INFO: eval_batch_step : [0, 400] [2024/12/11 13:35:07] ppocr INFO: infer_img : doc/imgs_en/img_10.jpg [2024/12/11 13:35:07] ppocr INFO: log_smooth_window : 20 [2024/12/11 13:35:07] ppocr INFO: pretrained_model : ./work/pretrain_models/ch_PP-OCRv3_det_distill_train/student [2024/12/11 13:35:07] ppocr INFO: print_batch_step : 1 [2024/12/11 13:35:07] ppocr INFO: save_epoch_step : 100 [2024/12/11 13:35:07] ppocr INFO: save_inference_dir : None [2024/12/11 13:35:07] ppocr INFO: save_model_dir : ./work/output/ch_PP-OCR_V3_det/ [2024/12/11 13:35:07] ppocr INFO: save_res_path : ./work/checkpoints/det_db/predicts_db.txt [2024/12/11 13:35:07] ppocr INFO: use_gpu : False [2024/12/11 13:35:07] ppocr INFO: use_visualdl : False [2024/12/11 13:35:07] ppocr INFO: Loss : [2024/12/11 13:35:07] ppocr INFO: alpha : 5 [2024/12/11 13:35:07] ppocr INFO: balance_loss : True [2024/12/11 13:35:07] ppocr INFO: beta : 10 [2024/12/11 13:35:07] ppocr INFO: main_loss_type : DiceLoss [2024/12/11 13:35:07] ppocr INFO: name : DBLoss [2024/12/11 13:35:07] ppocr INFO: ohem_ratio : 3 [2024/12/11 13:35:07] ppocr INFO: Metric : [2024/12/11 13:35:07] ppocr INFO: main_indicator : hmean [2024/12/11 13:35:07] ppocr INFO: name : DetMetric [2024/12/11 13:35:07] ppocr INFO: Optimizer : [2024/12/11 13:35:07] ppocr INFO: beta1 : 0.9 [2024/12/11 13:35:07] ppocr INFO: beta2 : 0.999 [2024/12/11 13:35:07] ppocr INFO: lr : [2024/12/11 13:35:07] ppocr INFO: learning_rate : 0.0005 [2024/12/11 13:35:07] ppocr INFO: name : Cosine [2024/12/11 13:35:07] ppocr INFO: warmup_epoch : 2 [2024/12/11 13:35:07] ppocr INFO: name : Adam [2024/12/11 13:35:07] ppocr INFO: regularizer : [2024/12/11 13:35:07] ppocr INFO: factor : 5e-05 [2024/12/11 13:35:07] ppocr INFO: name : L2 [2024/12/11 13:35:07] ppocr INFO: PostProcess : [2024/12/11 13:35:07] ppocr INFO: box_thresh : 0.6 [2024/12/11 13:35:07] ppocr INFO: max_candidates : 1000 [2024/12/11 13:35:07] ppocr INFO: name : DBPostProcess [2024/12/11 13:35:07] ppocr INFO: thresh : 0.3 [2024/12/11 13:35:07] ppocr INFO: unclip_ratio : 1.5 [2024/12/11 13:35:07] ppocr INFO: Train : [2024/12/11 13:35:07] ppocr INFO: dataset : [2024/12/11 13:35:07] ppocr INFO: data_dir : ./work/train_data/ [2024/12/11 13:35:07] ppocr INFO: label_file_list : ['./work/train_data/img/Label.txt'] [2024/12/11 13:35:07] ppocr INFO: name : SimpleDataSet [2024/12/11 13:35:07] ppocr INFO: ratio_list : [1.0] [2024/12/11 13:35:07] ppocr INFO: transforms : [2024/12/11 13:35:07] ppocr INFO: DecodeImage : [2024/12/11 13:35:07] ppocr INFO: channel_first : False [2024/12/11 13:35:07] ppocr INFO: img_mode : BGR [2024/12/11 13:35:07] ppocr INFO: DetLabelEncode : None [2024/12/11 13:35:07] ppocr INFO: IaaAugment : [2024/12/11 13:35:07] ppocr INFO: augmenter_args : [2024/12/11 13:35:07] ppocr INFO: args : [2024/12/11 13:35:07] ppocr INFO: p : 0.5 [2024/12/11 13:35:07] ppocr INFO: type : Fliplr [2024/12/11 13:35:07] ppocr INFO: args : [2024/12/11 13:35:07] ppocr INFO: rotate : [-10, 10] [2024/12/11 13:35:07] ppocr INFO: type : Affine [2024/12/11 13:35:07] ppocr INFO: args : [2024/12/11 13:35:07] ppocr INFO: size : [0.5, 3] [2024/12/11 13:35:07] ppocr INFO: type : Resize [2024/12/11 13:35:07] ppocr INFO: EastRandomCropData : [2024/12/11 13:35:07] ppocr INFO: keep_ratio : True [2024/12/11 13:35:07] ppocr INFO: max_tries : 50 [2024/12/11 13:35:07] ppocr INFO: size : [960, 960] [2024/12/11 13:35:07] ppocr INFO: MakeBorderMap : [2024/12/11 13:35:07] ppocr INFO: shrink_ratio : 0.4 [2024/12/11 13:35:07] ppocr INFO: thresh_max : 0.7 [2024/12/11 13:35:07] ppocr INFO: thresh_min : 0.3 [2024/12/11 13:35:07] ppocr INFO: MakeShrinkMap : [2024/12/11 13:35:07] ppocr INFO: min_text_size : 8 [2024/12/11 13:35:07] ppocr INFO: shrink_ratio : 0.4 [2024/12/11 13:35:07] ppocr INFO: NormalizeImage : [2024/12/11 13:35:07] ppocr INFO: mean : [0.485, 0.456, 0.406] [2024/12/11 13:35:07] ppocr INFO: order : hwc [2024/12/11 13:35:07] ppocr INFO: scale : 1./255. [2024/12/11 13:35:07] ppocr INFO: std : [0.229, 0.224, 0.225] [2024/12/11 13:35:07] ppocr INFO: ToCHWImage : None [2024/12/11 13:35:07] ppocr INFO: KeepKeys : [2024/12/11 13:35:07] ppocr INFO: keep_keys : ['image', 'threshold_map', 'threshold_mask', 'shrink_map', 'shrink_mask'] [2024/12/11 13:35:07] ppocr INFO: loader : [2024/12/11 13:35:07] ppocr INFO: batch_size_per_card : 8 [2024/12/11 13:35:07] ppocr INFO: drop_last : False [2024/12/11 13:35:07] ppocr INFO: num_workers : 4 [2024/12/11 13:35:07] ppocr INFO: shuffle : True [2024/12/11 13:35:07] ppocr INFO: profiler_options : None [2024/12/11 13:35:07] ppocr INFO: train with paddle 2.6.2 and device Place(cpu) [2024/12/11 13:35:07] ppocr INFO: Initialize indexs of datasets:['./work/train_data/img/Label.txt'] [2024/12/11 13:35:07] ppocr INFO: Initialize indexs of datasets:['./work/train_data/img/Label.txt'] [2024/12/11 13:35:08] ppocr INFO: train dataloader has 2 iters [2024/12/11 13:35:08] ppocr INFO: valid dataloader has 9 iters [2024/12/11 13:35:08] ppocr INFO: load pretrain successful from ./work/pretrain_models/ch_PP-OCRv3_det_distill_train/student [2024/12/11 13:35:08] ppocr INFO: During the training process, after the 0th iteration, an evaluation is run every 400 iterations [2024/12/11 13:35:55] ppocr INFO: epoch: [1/500], global_step: 1, lr: 0.000000, loss: 2.071318, loss_shrink_maps: 1.280234, loss_threshold_maps: 0.534262, loss_binary_maps: 0.256823, loss_cbn: 0.000000, avg_reader_cost: 2.17457 s, avg_batch_cost: 46.90998 s, avg_samples: 8.0, ips: 0.17054 samples/s, eta: 13:01:03, [2024/12/11 13:35:55] ppocr INFO: save model in ./work/output/ch_PP-OCR_V3_det/latest [2024/12/11 13:36:41] ppocr INFO: epoch: [2/500], global_step: 2, lr: 0.000062, loss: 2.127443, loss_shrink_maps: 1.316309, loss_threshold_maps: 0.547244, loss_binary_maps: 0.263890, loss_cbn: 0.000000, avg_reader_cost: 2.32840 s, avg_batch_cost: 46.07263 s, avg_samples: 8.0, ips: 0.17364 samples/s, eta: 12:52:31, [2024/12/11 13:36:41] ppocr INFO: save model in ./work/output/ch_PP-OCR_V3_det/latest [2024/12/11 13:37:26] ppocr INFO: epoch: [3/500], global_step: 3, lr: 0.000125, loss: 2.183568, loss_shrink_maps: 1.352385, loss_threshold_maps: 0.549175, loss_binary_maps: 0.270958, loss_cbn: 0.000000, avg_reader_cost: 2.12764 s, avg_batch_cost: 45.30531 s, avg_samples: 8.0, ips: 0.17658 samples/s, eta: 12:44:25,
# 导出检测模型
python tools/export_model.py -c configs/det/ch_PP-OCRv3/ch_PP-OCRv3_det_student_cimc.yml -o Global.pretrained_model="./work/output/ch_PP-OCR_V3_det/best_accuracy" Global.save_inference_dir="./work/inference_model/ch_PP-OCR_V3_det/"
# 执行预测
python tools/infer/predict_system.py --image_dir="./work/train_data/img/1.jpg" --det_model_dir="./work/inference_model/ch_PP-OCR_V3_det" --det_limit_side_len=48 --det_limit_type='min' --det_db_unclip_ratio=2.5 --rec_model_dir="./work/model/ch_PP-OCRv3_rec_infer" --rec_image_shape="3, 48, 320" --draw_img_save_dir=./work/det_rec_infer/ --use_space_char=False --use_angle_cls=False --use_gpu=false
识别效果：