app.py

import os
import glob
from itertools import groupby

import cv2
import numpy as np
import gradio as gr
import tensorflow as tf


def get_sample_images():
    list_ = glob.glob(os.path.join(os.path.dirname(__file__), 'samples/*.jpg'))
    return [[i] for i in list_]


def inference(image):
    # load model
    demo = TFliteDemo(os.path.join(os.path.dirname(__file__), 'model.tflite'))
    # load image
    image = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
    image = center_fit(image, 128, 64, top_left=True)
    image = np.reshape(image, (1, *image.shape, 1)).astype(np.uint8)
    # inference
    pred = demo.inference(image)
    # decode
    dict = load_dict(os.path.join(os.path.dirname(__file__), 'label.names'))
    res = decode_label(pred, dict)
    return res


class TFliteDemo:
    def __init__(self, model_path):
        self.interpreter = tf.lite.Interpreter(model_path=model_path)
        self.interpreter.allocate_tensors()
        self.input_details = self.interpreter.get_input_details()
        self.output_details = self.interpreter.get_output_details()
    
    def inference(self, x):
        self.interpreter.set_tensor(self.input_details[0]['index'], x)
        self.interpreter.invoke()
        return self.interpreter.get_tensor(self.output_details[0]['index'])


def center_fit(img, w, h, inter=cv2.INTER_NEAREST, top_left=True):
    # get img shape
    img_h, img_w = img.shape[:2]
    # get ratio
    ratio = min(w / img_w, h / img_h)

    if len(img.shape) == 3:
        inter = cv2.INTER_AREA
    # resize img
    img = cv2.resize(img, (int(img_w * ratio), int(img_h * ratio)), interpolation=inter)
    # get new img shape
    img_h, img_w = img.shape[:2]
    # get start point
    start_w = (w - img_w) // 2
    start_h = (h - img_h) // 2

    if top_left:
        start_w = 0
        start_h = 0

    if len(img.shape) == 2:
        # create new img
        new_img = np.zeros((h, w), dtype=np.uint8)
        new_img[start_h:start_h+img_h, start_w:start_w+img_w] = img
    else:
        new_img = np.zeros((h, w, 3), dtype=np.uint8)
        new_img[start_h:start_h+img_h, start_w:start_w+img_w, :] = img

    return new_img


def load_dict(dict_path='label.names'):
    with open(dict_path, 'r', encoding='utf-8') as f:
        dict = f.read().splitlines()
    dict = {i: dict[i] for i in range(len(dict))}
    return dict


def decode_label(mat, chars) -> str:
    # mat is the output of model
    # get char indices along best path
    best_path_indices = np.argmax(mat[0], axis=-1)
    # collapse best path (using itertools.groupby), map to chars, join char list to string
    best_chars_collapsed = [chars[k] for k, _ in groupby(best_path_indices) if k != len(chars)]
    res = ''.join(best_chars_collapsed)
    # remove space and '_'
    res = res.replace(' ', '').replace('_', '')
    return res

_TITLE = '''South Korean License Plate Recognition'''
_DESCRIPTION = '''
<div>
<p style="text-align: center; font-size: 1.3em">This is a demo of South Korean License Plate Recognition.
<a style="display:inline-block; margin-left: .5em" href='https://github.com/noahzhy/KR_LPR_TF/'><img src='https://img.shields.io/github/stars/noahzhy/KR_LPR_TF?style=social' /></a>
</p>
</div>
'''
interface = gr.Interface(
    fn=inference,
    inputs="image",
    outputs="text",
    title=_TITLE,
    description=_DESCRIPTION,
    examples=get_sample_images(),
)
interface.launch()