Доработки по скрипту для вырезания VIN из СТС.

.gitignore

@@ -13,6 +13,7 @@ output/
 train_data/
 pretrained_models/
 log/
+tags
 *.DS_Store
 *.vs
 *.user

configs/rec_custom.yml

@@ -1,8 +1,41 @@
 Global:
+  model_name: PP-OCRv5_server_rec # To use static model for inference.
+  debug: false
   use_gpu: false
+  epoch_num: 75
+  log_smooth_window: 20
+  print_batch_step: 10
+  save_model_dir: output/PP-OCRv5_server_rec
+  save_epoch_step: 1
+  eval_batch_step: [0, 2000]
+  cal_metric_during_train: true
+  calc_epoch_interval: 1
+  pretrained_model: https://paddle-model-ecology.bj.bcebos.com/paddlex/official_pretrained_model/PP-OCRv5_server_rec_pretrained.pdparams
+  checkpoints:
+  save_inference_dir:
+  use_visualdl: false
+  infer_img: train_data/images/1C4HJXEN3MW645094.jpg
   character_dict_path: train_data/dict.txt
-  save_model_dir: ./output/
+  max_text_length: &max_text_length 19
-  pretrained_model: pretrained_models/PP-OCRv5_server_rec_pretrained.pdparams
+  infer_mode: false
+  use_space_char: true
+  distributed: true
+  save_res_path: output/rec/predicts_ppocrv5.txt
+  d2s_train_image_shape: [3, 48, 320]
+
+
+Optimizer:
+  name: Adam
+  beta1: 0.9
+  beta2: 0.999
+  lr:
+    name: Cosine
+    learning_rate: 0.0005
+    warmup_epoch: 1
+  regularizer:
+    name: L2
+    factor: 3.0e-05
+
 
 Architecture:
   model_type: rec
@@ -26,7 +59,13 @@ Architecture:
             fc_decay: 0.00001
       - NRTRHead:
           nrtr_dim: 384
-          max_text_length: 17
+          max_text_length: *max_text_length
+
+Loss:
+  name: MultiLoss
+  loss_config_list:
+    - CTCLoss:
+    - NRTRLoss:
 
 PostProcess:  
   name: CTCLabelDecode
@@ -37,18 +76,41 @@ Metric:
 
 Train:
   dataset:
-    name: SimpleDataSet
+    name: MultiScaleDataSet
-    data_dir: train_data/
+    ds_width: false
+    data_dir: train_data/images
+    ext_op_transform_idx: 1
     label_file_list:
     - train_data/train.txt
+    transforms:
+    - DecodeImage:
+        img_mode: BGR
+        channel_first: false
+    - MultiLabelEncode:
+        gtc_encode: NRTRLabelEncode
+    - KeepKeys:
+        keep_keys:
+        - image
+        - label_ctc
+        - label_gtc
+        - length
+        - valid_ratio
+  sampler:
+    name: MultiScaleSampler
+    scales: [[320, 32], [320, 48], [320, 64]]
+    first_bs: &bs 128
+    fix_bs: false
+    divided_factor: [8, 16] # w, h
+    is_training: True
   loader:
-    batch_size_per_card: 16  # Уменьшите для экономии памяти
+    shuffle: true
-    num_workers: 2
+    batch_size_per_card: *bs
-
+    drop_last: true
+    num_workers: 16
 Eval:
   dataset:
     name: SimpleDataSet
-    data_dir: train_data/
+    data_dir: train_data/images
     label_file_list:
     - train_data/val.txt
     transforms:
@@ -57,7 +119,6 @@ Eval:
         channel_first: false
     - MultiLabelEncode:
         gtc_encode: NRTRLabelEncode
-        max_text_length: 17
     - RecResizeImg:
         image_shape: [3, 48, 320]
     - KeepKeys:

configs/rec_custom_kaggle.yml

@@ -0,0 +1,135 @@
+Global:
+  model_name: PP-OCRv5_server_rec # To use static model for inference.
+  debug: false
+  use_gpu: true
+  epoch_num: 75
+  log_smooth_window: 20
+  print_batch_step: 10
+  save_model_dir: /kaggle/output/PP-OCRv5_server_rec
+  save_epoch_step: 1
+  eval_batch_step: [0, 2000]
+  cal_metric_during_train: true
+  calc_epoch_interval: 1
+  pretrained_model: https://paddle-model-ecology.bj.bcebos.com/paddlex/official_pretrained_model/PP-OCRv5_server_rec_pretrained.pdparams
+  checkpoints:
+  save_inference_dir:
+  use_visualdl: false
+  infer_img: /kaggle/input/custom-ocr-dataset/images/1C4HJXEN3MW645094.jpg
+  character_dict_path: /kaggle/input/custom-ocr-dataset/dict.txt
+  max_text_length: &max_text_length 19
+  infer_mode: false
+  use_space_char: true
+  distributed: true
+  save_res_path: /kaggle/output/rec/predicts_ppocrv5.txt
+  d2s_train_image_shape: [3, 48, 320]
+
+
+Optimizer:
+  name: Adam
+  beta1: 0.9
+  beta2: 0.999
+  lr:
+    name: Cosine
+    learning_rate: 0.0005
+    warmup_epoch: 1
+  regularizer:
+    name: L2
+    factor: 3.0e-05
+
+
+Architecture:
+  model_type: rec
+  algorithm: SVTR_HGNet
+  Transform:
+  Backbone:
+    name: PPHGNetV2_B4
+    text_rec: True
+  Head:
+    name: MultiHead
+    head_list:
+      - CTCHead:
+          Neck:
+            name: svtr
+            dims: 120
+            depth: 2
+            hidden_dims: 120
+            kernel_size: [1, 3]
+            use_guide: True
+          Head:
+            fc_decay: 0.00001
+      - NRTRHead:
+          nrtr_dim: 384
+          max_text_length: *max_text_length
+
+Loss:
+  name: MultiLoss
+  loss_config_list:
+    - CTCLoss:
+    - NRTRLoss:
+
+PostProcess:  
+  name: CTCLabelDecode
+
+Metric:
+  name: RecMetric
+  main_indicator: acc
+
+Train:
+  dataset:
+    name: MultiScaleDataSet
+    ds_width: false
+    data_dir: /kaggle/input/custom-ocr-dataset/images
+    ext_op_transform_idx: 1
+    label_file_list:
+    - /kaggle/input/custom-ocr-dataset/train.txt
+    transforms:
+    - DecodeImage:
+        img_mode: BGR
+        channel_first: false
+    - MultiLabelEncode:
+        gtc_encode: NRTRLabelEncode
+    - KeepKeys:
+        keep_keys:
+        - image
+        - label_ctc
+        - label_gtc
+        - length
+        - valid_ratio
+  sampler:
+    name: MultiScaleSampler
+    scales: [[320, 32], [320, 48], [320, 64]]
+    first_bs: &bs 128
+    fix_bs: false
+    divided_factor: [8, 16] # w, h
+    is_training: True
+  loader:
+    shuffle: true
+    batch_size_per_card: *bs
+    drop_last: true
+    num_workers: 16
+Eval:
+  dataset:
+    name: SimpleDataSet
+    data_dir: /kaggle/input/custom-ocr-dataset/images
+    label_file_list:
+    - /kaggle/input/custom-ocr-dataset/val.txt
+    transforms:
+    - DecodeImage:
+        img_mode: BGR
+        channel_first: false
+    - MultiLabelEncode:
+        gtc_encode: NRTRLabelEncode
+    - RecResizeImg:
+        image_shape: [3, 48, 320]
+    - KeepKeys:
+        keep_keys:
+        - image
+        - label_ctc
+        - label_gtc
+        - length
+        - valid_ratio
+  loader:
+    shuffle: false
+    drop_last: false
+    batch_size_per_card: 128
+    num_workers: 4

requirements.txt

@@ -15,8 +15,8 @@ albumentations
 # to be compatible with albumentations
 albucore
 packaging
-# paddlepaddle-gpu
+paddlepaddle-gpu==2.5.2
-paddlepaddle
+# paddlepaddle
 paddleocr
 imgaug
 lmdb

scripts/dataset_validation.py

@@ -1,5 +1,7 @@
-import os
 import cv2
+import os
+import random
+from PIL import Image
 
 def rename_files(directory_path: str):
     extensions = set()
@@ -37,8 +39,51 @@ def check_symbols(dir_path: str):
             if any([char not in dict_chars for char in label]):
                 print(filename)
 
+def max_height(dir_path: str):
+    max_height = 0
+    max_filename = ''
+    for filename in os.listdir(dir_path):
+        im = Image.open(os.path.join(dir_path, filename))
+        if im.height > max_height:
+            max_height = im.height
+            max_filename = filename
+
+    print(max_filename, max_height)
+
+def resize_to_height(dir_path: str, target_height=48):
+    for filename in os.listdir(dir_path):
+        with Image.open(os.path.join(dir_path, filename)) as img:
+            width_percent = target_height / float(img.height)
+            new_width = int(float(img.width) * width_percent)
+            resized_img = img.resize((new_width, target_height), Image.LANZOS)
+            resized_img.save(os.path.join(dir_path, filename))
+
+def split_dataset(dir_path: str):
+    images_dir = os.path.join(dir_path, "images")
+    filenames = os.listdir(images_dir)
+    dataset_length = len(filenames)
+    random.shuffle(filenames)
+    train_ratio = 0.8
+    val_ratio = 0.2
+    train_files_len = round(dataset_length * train_ratio)
+    val_files_len = round(dataset_length * val_ratio)
+    train_files = filenames[:train_files_len]
+    val_files = filenames[train_files_len:train_files_len+val_files_len]
+    train_file_path = os.path.join(dir_path, "train.txt")
+    val_file_path = os.path.join(dir_path, "val.txt")
+    with open(train_file_path, "w") as train_file:
+        for filename in train_files:
+            label, _ = filename.split(".")
+            train_file.write(f"{filename}\t{label}\n")
+
+    with open(val_file_path, "w") as val_file:
+        for filename in val_files:
+            label, _ = filename.split(".")
+            val_file.write(f"{filename}\t{label}\n")
+
 # rename_files("train_data/images/")
 # check_images("train_data/images/")
 # check_labels("train_data/images/")
-check_symbols("train_data/")
+# check_symbols("train_data/")
-
+# max_height("train_data/images")
+split_dataset("train_data/")

scripts/detect_vin.py

@@ -0,0 +1,45 @@
+"""
+exec(open("scripts/detect_vin.py").read())
+"""
+
+from paddleocr import PaddleOCR
+import cv2
+import re
+
+ocr = PaddleOCR(lang='en', use_textline_orientation=True)
+
+def cut_vin(ocr, input_image, output_image):
+    """
+    Вырезать VIN из изображения.
+    """
+    
+    image = cv2.imread(input_image)
+
+    result = ocr.predict(image)
+    processed_image = result[0]["doc_preprocessor_res"]["output_img"]
+
+    vin_pattern = re.compile(r'^[A-HJ-NPR-Z0-9]{17}$')
+    found_vin = None
+
+    for text, bbox in zip(result[0]["rec_texts"], result[0]["rec_boxes"]):
+        if vin_pattern.match(text):
+            found_vin = text
+            break
+
+    if found_vin:
+        x_min, y_min = bbox[0], bbox[1]
+        x_max, y_max = bbox[2], bbox[3]
+
+        vin_region = processed_image[y_min:y_max, x_min:x_max]
+
+        cv2.imwrite(output_image, vin_region)
+
+    return found_vin
+
+vin = cut_vin(ocr, "input/image.jpg", "output/vin.jpg")
+
+if vin:
+    print(f"VIN найден: {vin}.")
+else:
+    print("VIN не обнаружен.")
+    

scripts/rec_validation.py

@@ -0,0 +1,27 @@
+import os
+from paddleocr import TextRecognition
+
+ocr = TextRecognition(
+    model_name="PP-OCRv5_server_rec"
+    # model_dir="output/PP-OCRv5_server_rec_vin"
+)
+
+with open("train_data/val.txt", "r") as label_file:
+    lines = label_file.readlines()
+    total = len(lines)
+    matches, mismatches = 0, 0
+    for idx, line in enumerate(lines, start=1):
+        file_name, label = line.split("\t")
+        label = label.strip()
+        path = os.path.join("train_data", os.path.join("images", file_name))
+        result = ocr.predict(path)
+        print(f"{idx}/{total} ", end="")
+        if result[0]["rec_text"] and result[0]["rec_text"] == label:
+            print(f"match {label}")
+            matches += 1
+        else:
+            print(f"mismatch {label}")
+            mismatches += 1
+
+print(f"{matches} matches of {total}")
+print(f"{mismatches} mismatches of {total}")

scripts/score_validation.py

@@ -0,0 +1,27 @@
+import os
+from paddleocr import PaddleOCR
+from typing import Optional
+
+ocr = PaddleOCR(rec_model_dir="output/vin_rec_inference")
+
+def get_best_accuracy(ocr: PaddleOCR, file_path: str) -> Optional[tuple[str, float]]:
+    result = ocr.predict(file_path)
+    texts_with_scores = [item for item in filter(
+        lambda x: len(x[0]) == 17,
+        zip(result[0]["rec_texts"], result[0]["rec_scores"])
+    )]
+    if texts_with_scores:
+        return max(
+            texts_with_scores,
+            key=lambda y: y[1]
+        )
+    else:
+        return None
+        
+
+with open("train_data/val.txt", "r") as label_file:
+    for line in label_file.readlines():
+        file_name, label = line.split("\t")
+        path = os.path.join("train_data", os.path.join("images", file_name))
+        print(file_name)
+        print(get_best_accuracy(ocr, path))