Empty result when use trained model DB++

[FahNos]

Hi master !
I am newbie, have just use PaddleOCR
When I use trained mode model in link https://paddleocr.bj.bcebos.com/dygraph_v2.1/en_det/det_r50_db%2B%2B_icdar15_train.tar
or pretrained model https://paddleocr.bj.bcebos.com/dygraph_v2.1/en_det/ResNet50_dcn_asf_synthtext_pretrained.pdparams
and command :
!python3 '/content/PaddleOCR/tools/infer/predict_det.py' \ --use_gpu=False \ --det_model_dir="/content/models/ResNet50_dcn_asf_synthtext_pretrained" \ --image_dir="/content/2.jpg"

The result I had is empty.

but When I use this code:
from paddleocr import PaddleOCR,draw_ocr
ocr = PaddleOCR( use_angle_cls=False,lang='en')
result = ocr.ocr(img_path, cls=False)
for line in result:
print(line)

the resule is good:

Can you help me explain this error and how to resolve it ?

Thank you so much

[GreatV]

Hi! Let’s analyze the issue and explore why you're getting empty results with your command-line inference using the trained model, while the PaddleOCR library approach works correctly.

Possible Causes and Solutions

1. Incorrect Model Directory Structure:

   • The --det_model_dir parameter should point to the directory containing the exported model files (inference.pdmodel, inference.pdiparams, and inference.pdiparams.info).
   • Ensure that the directory /content/models/ResNet50_dcn_asf_synthtext_pretrained contains these files. If not, you may need to extract the .tar file you downloaded or ensure that the correct model is used.

   Solution:
   Check the contents of /content/models/ResNet50_dcn_asf_synthtext_pretrained. If the files are missing, download and extract the model properly:

   tar -xvf det_r50_db++_icdar15_train.tar -C /content/models

2. Mismatch Between Model Type and Code:

   • The model you downloaded (det_r50_db++_icdar15_train) is a detection model. However, from your code, the results might suggest you're trying to perform OCR (text recognition) directly.
   • The predict_det.py script is specifically for text detection, not recognition. It will return empty results if the image lacks detectable text regions or if there’s a configuration mismatch.

   Solution:
   Use the appropriate script for recognition if your goal is OCR. Combine detection and recognition results using the full PaddleOCR pipeline.

   Alternatively, use:

   python3 /content/PaddleOCR/tools/infer/predict_system.py \
       --use_gpu=False \
       --det_model_dir="/content/models/det_r50_db++_icdar15_train" \
       --image_dir="/content/2.jpg"

   The predict_system.py script handles both detection and recognition seamlessly.

3. Incorrect Preprocessing or Parameters:

   • When using predict_det.py, ensure the preprocessing steps match the model's training setup. The det_db_thresh, det_db_box_thresh, and other parameters may need adjustment.
   • By default, the thresholds might not be optimal for your specific image.

   Solution:
   Try tuning the detection thresholds:

   python3 /content/PaddleOCR/tools/infer/predict_det.py \
       --use_gpu=False \
       --det_model_dir="/content/models/det_r50_db++_icdar15_train" \
       --image_dir="/content/2.jpg" \
       --det_db_thresh=0.3 \
       --det_db_box_thresh=0.5

4. Image Format or Path Issue:

   • Ensure that /content/2.jpg is a valid image file and in a format supported by the detection model.
   • If the file path or format is incorrect, the model will not process the image, resulting in empty output.

   Solution:
   Validate the image:

   from PIL import Image
   img = Image.open("/content/2.jpg")
   img.show()

   Ensure the path is correct and the image is suitable for text detection.

5. Differences Between PaddleOCR API and CLI Tools:

   • The PaddleOCR Python API (ocr = PaddleOCR(...)) internally handles both text detection and recognition, along with preprocessing. This is why it works well with minimal configuration.
   • However, the CLI tools (predict_det.py) require explicit setup for the specific task (detection or recognition).

   Solution:
   If the PaddleOCR API works better for your use case, consider using it directly in your Python script instead of the CLI tools.

---

Summary of Steps to Resolve:

1. Verify Model Directory:
   Ensure /content/models/ResNet50_dcn_asf_synthtext_pretrained contains inference.pdmodel, inference.pdiparams, and inference.pdiparams.info.

2. Use the Full Pipeline:
   Use predict_system.py if you want both detection and recognition:

   python3 /content/PaddleOCR/tools/infer/predict_system.py \
       --use_gpu=False \
       --det_model_dir="/content/models/det_r50_db++_icdar15_train" \
       --image_dir="/content/2.jpg"

3. Tune Parameters:
   Adjust detection thresholds (det_db_thresh, det_db_box_thresh) if detection results are still empty.

4. Validate the Image:
   Ensure the image path and format are correct.

By applying these fixes, you should be able to resolve the issue and obtain results similar to those from the PaddleOCR library. Let me know if you need further assistance!

---

Response generated by feifei-bot | chatgpt-4o-latest

[FahNos]

I use it command:
!python3 '/content/PaddleOCR/tools/infer/predict_system.py' \ --use_gpu=True \ --det_algorithm="SAST" \ --det_model_dir="/content/models/det_r50_vd_sast_icdar15_v2.0_train" \ --rec_algorithm="SRN" \ --rec_model_dir="/content/inference/rec_srn_vnese" \ --rec_image_shape="1, 64, 256" \ --image_dir="/content/inference/inference_result/1.jpg" \ --rec_char_type="ch" \ --drop_score=0.7 \ --rec_char_dict_path="/content/vi_vietnam.txt"

But have error:
predict_system.py: error: unrecognized arguments: --rec_char_type=ch

Can you help me resolve it ?