leopits / ocr-dot-matrix-text

OCR to detect and recognize dot-matrix text written with inkjet-printed on medical PVC bag

OCR-Dotted-Matrix

OCR to detect and recognize dot-matrix text written with inkjet-printed on medical PVC bag

Example images:

TEXT DETECTION wiht CRAFT (Character-Region Awareness For Text detection)

The code pre-processes images with the OpenCV function to improve text detection with CRAFT with (https://github.com/clovaai/CRAFT-pytorch/blob/master/README.md#craft-character-region-awareness-for-text-detection) The weights of pre-train network are available on this link https://drive.google.com/file/d/1Jk4eGD7crsqCCg9C9VjCLkMN3ze8kutZ/view. The recognize label is a string of the text, so the CRAFT parameters are set to find a unique block of text. it is possible to change --text_threshold,--low_text ,--link_threshold to have different detection results, but it is necessary to modify the label and recognition method after.

Craft results:

TEXT RECOGNITION with TESSERACT

The code extract the area around text on original image and fix the text oriention.

The cropped image:

Morphology Transformations (OpenCV function) and rescaling of chars with different parameters are applied to the cropped image.

Pre-process cropped image:

I use Tesseract OCR engine (https://tesseract-ocr.github.io/) with default page segmentation , the experiments show the LCDDot_FT_500.traineddata performs the best results in this case. Two methods are used to control the label:

• SequenceMatcher is a class available in python module named difflib. It can be used for comparing pairs of input sequences. With the function ratio( ) returns the similarity score ( float in [0,1] ) between input strings. It sums the sizes of all matched sequences returned by function.
• Regular expression is a class available in python module named re. The function re.match() checks for a match only at the beginning of the string.

Saving all result in json file:

        {
            "Name_original_file": "A_0.png",
            "Name_preprocess": "_preprocess_150.jpg",
            "check_label": "LOTTO:L21X45SCAD.:10-2023",
            "tesseract_LCDDot_FT_500_psm3_result": "LOTTO:L21X45SCAD.:10-2023",
            "LCDDot_FT_500_psm3_sequence_matcher_ratio_result": 1.0,
            "LCDDot_FT_500_psm3_bool_re_result": true
        }
    ],
    [
        {
            "Name_original_file": "A_0.png",
            "Name_preprocess": "_preprocess_160.jpg",
            "check_label": "LOTTO:L21X45SCAD.:10-2023",
            "tesseract_LCDDot_FT_500_psm3_result": "LOTTO:L21X4SCAD.:1625555",
            "LCDDot_FT_500_psm3_sequence_matcher_ratio_result": 0.78,
            "LCDDot_FT_500_psm3_bool_re_result": false
        }
    ],
    [
        {
            "Name_original_file": "A_0.png",
            "Name_preprocess": "_preprocess_170.jpg",
            "check_label": "LOTTO:L21X45SCAD.:10-2023",
            "tesseract_LCDDot_FT_500_psm3_result": "LOTTO:L21X45SCAD.:10-2023",
            "LCDDot_FT_500_psm3_sequence_matcher_ratio_result": 1.0,
            "LCDDot_FT_500_psm3_bool_re_result": true
        }

Getting started

Install dependencies

Requirements

• PyTorch>=1.9.0
• torchvision>=0.2.2
• opencv-python>=4.5.2

conda env create -f environment.yml

Run script

python Test_Image.py  --image [folder path to test images]  --folder_res [folder path to save result images] --label [string label to check]

Arguments

• --image: folder path to test images

• --label: string label to check

• --folder_res: folder path to save result images

• --trained_model: pretrained model

• --text_threshold: text confidence threshold

• --low_text: text low-bound score

• --link_threshold: link confidence threshold

• --cuda: use cuda for inference (default:True)

• --canvas_size: max image size for inference

• --mag_ratio: image magnification ratio

• --poly: enable polygon type result

• --show_time: show processing time

• --test_folder: folder path to input images

• --refine: use link refiner for sentense-level dataset

• --refiner_model: pretrained refiner model