Making a Tesseract model for my license plates
Requirements
First download the following repository:git clone https://github.com/tesseract-ocr/tesstrain.git
Navigate to the folder: cd tesstrain
Compile the application:make tesseract-langdata
Make a new to save the default language model to: mkdir -p usr/share/tessdata
Now download the model: wget -P usr/share/tessdata https://github.com/tesseract-ocr/tessdata_best/raw/main/eng.traineddata
Making the dataset
I made pictures of random cars outside for my other model. When you have a decently sized dataset (I used 279 images) make sure to extract the license plates from the images using the data from Label Studio. This is a python script I used:
extract_licenseplates.pyimport os import cv2 # Define paths labels_path = 'dataset/licensePlate/yolo/labels/' images_path = 'dataset/licensePlate/yolo/images' output_path = 'licenseplates/' # Make sure output directory exists os.makedirs(output_path, exist_ok=True) def extract_license_plate(image_name): # Load image img_path = os.path.join(images_path, image_name) image = cv2.imread(img_path) height, width, _ = image.shape # Corresponding label file label_file_name = f"{os.path.splitext(image_name)[0]}.txt" label_file_path = os.path.join(labels_path, label_file_name) if not os.path.exists(label_file_path): print(f"No label found for {image_name}") return with open(label_file_path) as f: lines = f.readlines() for line in lines: parts = line.strip().split() class_id = int(parts[0]) if class_id == 1: x_center, y_center, w, h= map(float ,parts [1:] ) x_min=int((x_center-w /2)*width ) y_min=int((y_center-h /2)* height ) x_max=int((x_center+ w /2)*width ) y_max=int((y_center+h /2 )*height ) cropped_img=image[y_min:y_max,x_min:x_max] # Save the cropped image as a PNG file png_output_file_name=os.path.join(output_path,f"lp_{os.path.splitext(image_name)[0]}.png") cv2.imwrite(png_output_file_name,cropped_img) print(f"Cropped image saved as {png_output_file_name}") for image_filename in os.listdir(images_path): if image_filename.endswith(('.jpg','.jpeg','.png')): extract_license_plate(image_filename) print("License plate extraction complete.")
NOTE: The images have to be in PNG format. If they aren’t, run this.mogrify -format png *.*
Process the dataset
This step will involve quite a bit of work, sadly. You will need to create a file, a .tg.txt file with the same name as the image. Here is a python script using gradio to make this a bit easier. Please note the image_folder. Point that folder to your dataset.
app.pyimport gradio as gr from PIL import Image import os image_folder = "licenseplates" image_files = [os.path.join(image_folder, f) for f in os.listdir(image_folder) if os.path.isfile(os.path.join(image_folder, f))] current_image_index = 0 def load_image_and_text(evt: gr.SelectData): global current_image_index, image_files, corrected_output image_path = os.path.join(image_folder, evt.value["image"]["orig_name"]) current_image_index = image_files.index(image_path) text = read_text(image_path) return image_path, text # Create a Gradio Interface with gr.Blocks() as app: # Function to save the corrected text def save_corrected_text_gr(filename, text): original_filename = filename.rsplit("/", 1)[-1].rsplit(".", 1)[0] with open(f"{image_folder}/{original_filename}.gt.txt", "w") as text_file: text_file.write(text) def read_text(image_file_path): # Extract the base filename without the directory path and extension base_filename = os.path.basename(image_file_path).rsplit(".", 1)[0] text_file_path = f"{image_folder}/{base_filename}.gt.txt" # Check if the corresponding .gt.txt file exists if os.path.exists(text_file_path): # Read the text from the file with open(text_file_path, "r") as file: text = file.read() else: # If the file does not exist, set text to an empty string text = "" return text # Update the next_image and previous_next functions to refresh the image component # Update the next_image and previous_image functions to use read_text def next_image(): global current_image_index, image_display, image_files, corrected_output if current_image_index < len(image_files) - 1: current_image_index += 1 n_image = image_files[current_image_index] # Read text from corresponding text file text = read_text(n_image) return (n_image, text) def previous_image(): global current_image_index, image_display, image_files, corrected_output if current_image_index > 0: current_image_index -= 1 p_image = image_files[current_image_index] # Read text from corresponding text file text = read_text(p_image) return (p_image, text) with gr.Row(): with gr.Column(scale=1): image_list = gr.Gallery(label="Image Gallery", value=image_files, allow_preview=False) with gr.Column(scale=4): current_image_file = gr.Textbox(image_files[current_image_index]) image_display = gr.Image(current_image_file.value) current_image_file.change(fn=lambda o: o, inputs=[current_image_file], outputs=[image_display]) corrected_output = gr.Textbox(label="Corrected Text", interactive=True) # Create the buttons to go to the next and previous image next_button = gr.Button("Next Image").click(fn=next_image, inputs=[], outputs=[current_image_file, corrected_output]) previous_button = gr.Button("Previous Image").click(fn=previous_image, inputs=[], outputs=[current_image_file, corrected_output]) # Create the button to save the corrected text save_button = gr.Button("Save Corrected Text").click( fn=save_corrected_text_gr, inputs=[current_image_file, corrected_output], outputs=[] ) image_list.select(load_image_and_text, outputs=[current_image_file, corrected_output]) # Launch the interface app.launch()
It will look like this.
Train the model
First navigate to the tesstrain folder you cloned before:cd tesstrain
Now copy your dataset to the required folder:cp -r {LOCATION_OF_DATASET}/* data/licenseplates-ground-truth
Train the model:make training MODEL_NAME=licenseplates START_MODEL=eng FINETUNE_TYPE=Impact
Move the new model to the model folder:sudo cp data/licenseplates.traineddata /usr/share/tessdata
Using the model
To use the new model:tesseract data/licenseplates-ground-truth/lp_5c3f36e7-20240601_194130.png output -l licenseplates
NOTE: The image you are going to run this on need to be processed by this script:
deskew.pyimport cv2 import numpy as np from scipy.ndimage import interpolation as inter import os def rotate_image(image, angle): (h, w) = image.shape[: 2] center = (w // 2, h // 2) M = cv2.getRotationMatrix2D(center, angle, 1.0) corrected = cv2.warpAffine(image, M, (w, h), flags = cv2.INTER_CUBIC, \ borderMode = cv2.BORDER_REPLICATE) return corrected def determine_score(arr): histogram = np.sum(arr, axis = 2, dtype = float) score = np.sum((histogram[..., 1 :] - histogram[..., : -1]) ** 2, \ axis = 1, dtype = float) return score def correct_skew(image, delta = 0.1, limit = 20): thresh = cv2.threshold(image, 0, 255, cv2.THRESH_BINARY_INV | cv2.THRESH_OTSU)[1] angles = np.arange(-limit, limit + delta, delta) img_stack = np.stack([rotate_image(thresh, angle) for angle in angles], axis = 0) scores = determine_score(img_stack) best_angle = angles[np.argmax(scores)] corrected = rotate_image(image, best_angle) return best_angle, corrected def correct_image(image): # treshold color outfile = cv2.threshold(corrected, 66, 255, cv2.THRESH_BINARY_INV | cv2.THRESH_OTSU)[1] # invert image return (255-outfile) image_folder = 'lp' output_folder = 'lp2' if not os.path.exists(output_folder): os.makedirs(output_folder) for filename in os.listdir(image_folder): if filename.endswith('.png'): image = cv2.imread(os.path.join(image_folder, filename), 0) angle, corrected = correct_skew(image) output_filename = os.path.join(output_folder, filename) cv2.imwrite(output_filename + ".png", correct_image(corrected))