recogPlate_all.py

def emptydir(dirname):  #清空資料夾
    if os.path.isdir(dirname):  #資料夾存在就刪除
        shutil.rmtree(dirname)
        sleep(2)  #需延遲,否則會出錯
    os.mkdir(dirname)  #建立資料夾

from keras.models import load_model
from PIL import Image
import numpy as np
import cv2
import shutil, os
from time import sleep 
import glob

labels = ['0','1','2','3','4','5','6','7','8','9','A','B','C','D','E','F','G','H','J','K','L','M','N','P','Q','R','S','T','U','V','W','X','Y','Z']  #標籤值
dirname = 'recogdata'
model = load_model("carplate_model.hdf5")  #讀取模型
myfiles = glob.glob('predictPlate/*.jpg')
for imgname in myfiles:
    #擷取車牌
    emptydir(dirname)
    img = cv2.imread(imgname)
    detector = cv2.CascadeClassifier('haar_carplate.xml')
    signs = detector.detectMultiScale(img, scaleFactor=1.01, minNeighbors=16, minSize=(20, 20))
    if len(signs) > 0 :
        for (x, y, w, h) in signs:          
            image1 = Image.open(imgname)
            image2 = image1.crop((x, y, x+w, y+h))
            image3 = image2.resize((140, 40), Image.ANTIALIAS)
            image3.save('tem.jpg')
            image4 = cv2.imread('tem.jpg')
            gray = cv2.cvtColor(image4, cv2.COLOR_RGB2GRAY)
            _, img_thre = cv2.threshold(gray, 100, 255, cv2.THRESH_BINARY)
            cv2.imwrite('tem.jpg', img_thre)
        #分割文字
        img_tem = cv2.imread('tem.jpg')
        gray = cv2.cvtColor(img_tem, cv2.COLOR_BGR2GRAY)
        _,thresh = cv2.threshold(gray, 127, 255, cv2.THRESH_BINARY_INV)  #轉為黑白
        contours1 = cv2.findContours(thresh.copy(), cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)  #尋找輪廓
        contours = contours1[1]  #取得輪廓
        letter_image_regions = []  #文字圖形串列
        for contour in contours:  #依序處理輪廓
            (x, y, w, h) = cv2.boundingRect(contour)  #單一輪廓資料
            letter_image_regions.append((x, y, w, h))  #輪廓資料加入串列
        letter_image_regions = sorted(letter_image_regions, key=lambda x: x[0])  #按X坐標排序
        #存檔
        i=1
        for letter_bounding_box in letter_image_regions:  #依序處理輪廓資料
            x, y, w, h = letter_bounding_box
            if w>=5 and h>32 and h<40:  #長度>6且高度在30-48才是文字
                letter_image = gray[y:y+h, x:x+w]  #擷取圖形
                letter_image = cv2.resize(letter_image, (18, 38))
                cv2.imwrite(dirname + '/{}.jpg'.format(i), letter_image)  #存檔
                i += 1
        #辨識車牌
        datan = 0  #車牌文字數,即檔案數
        for fname in os.listdir(dirname):
            if os.path.isfile(os.path.join(dirname, fname)):
                datan += 1
        tem_data = []
        for index in range(1, (datan+1)):  #讀取預測資料
            tem_data.append((np.array(Image.open("recogdata/" + str(index) +".jpg")))/255.0)
        real_data = np.stack(tem_data)
        real_data1 = np.expand_dims(real_data, axis=3)  #轉換維度符合需求->(7,38,18,1)
        predictions = model.predict_classes(real_data1)  #預測資料
        result = []
        for i in range(len(predictions)):  #顯示結果
            result.append(labels[int(predictions[i])])
        print(imgname + ' --> ' + ''.join(result)) 
    else:
        print('無法擷取車牌！')
    os.remove('tem.jpg')