pip install tensorflow==2.2.0
pip install tensorflow-gpu==2.2.0

pip install imageai

>> MainFold     >> WriteXML.py
                >> part.py
                >> Class.py
                >> Partition.py
                >> AddBackground.py
                >> DetectSelect.py
                >> ClassifySelect.py
                >> Train.py
                >> InceptionResNetV2.py
                >> ClassificationV2.py

                >> img (train 解壓縮)
                >> json (train 解壓縮)
                >> img_public (public 解壓縮)
                >> xml
                >> part_img
                >> Class
                >> AfterData
                >> target_data
                >> DetectTrain
                >> ClassifyTrain

\n " + data['imagePath'] + " \n \n") write=file.write(" " + str(the_width) + " \n " + str(the_heigh) + " \n") write=file.write(" " + str(the_depth) + " \n \n") for j in range(0, len(data['shapes'])): x_list, y_list = [], [] if (data['shapes'][j]['group_id'] == 1):#取出中文字元 for points in range(0, 4): x_list.append(data['shapes'][j]['points'][points][0]) y_list.append(data['shapes'][j]['points'][points][1]) xmax, xmin = max(x_list), min(x_list) ymax, ymin = max(y_list), min(y_list) write=file.write("\n") elif (data['shapes'][j]['group_id'] == 2):#取出英文數字字串 for points in range(0, 4): x_list.append(data['shapes'][j]['points'][points][0]) y_list.append(data['shapes'][j]['points'][points][1]) xmax, xmin = max(x_list), min(x_list) ymax, ymin = max(y_list), min(y_list) write = file.write("\n") del x_list, y_list write = file.write("") file.close() os.rename("xml/foruse.txt", "xml/" + data['imagePath'][:-4] + ".xml")">

import os
import shutil
import json
import cv2

json_data = os.listdir('json/')
num = 0
for i in json_data:   
    num += 1
    print(num)
    with open("json/" + i, encoding="utf-8") as f:
        
        data = json.load(f) #讀取json
        
        img = cv2.imread("img/" + data['imagePath'])
        the_width = data['imageWidth']
        the_heigh = data['imageHeight']
        the_depth = img.shape[2]
        
        #XML撰寫準備--------------------------------------
        space = "space.txt"
        foruse = "xml/foruse.txt"
        shutil.copy(space, foruse)
        file = open("xml/foruse.txt", mode = "w")
        
        write=file.write("
   
    \n
    
     "
    
    + data['imagePath'] + "\n
   
    \n"
   )
        write=file.write("
   
    "
    + str(the_width) + "\n
   
    "
    + str(the_heigh) + "\n")
        write=file.write("
   
    "
    + str(the_depth) + "\n\n")
        
        for j in range(0, len(data['shapes'])): 
            
            x_list, y_list = [], []
            
            if (data['shapes'][j]['group_id'] == 1):#取出中文字元
                
                for points in range(0, 4):
                    x_list.append(data['shapes'][j]['points'][points][0])
                    y_list.append(data['shapes'][j]['points'][points][1])         
                    xmax, xmin = max(x_list), min(x_list)
                    ymax, ymin = max(y_list), min(y_list)
                
                write=file.write("\n")
            
            elif (data['shapes'][j]['group_id'] == 2):#取出英文數字字串
                
                for points in range(0, 4):
                    x_list.append(data['shapes'][j]['points'][points][0])
                    y_list.append(data['shapes'][j]['points'][points][1])         
                    xmax, xmin = max(x_list), min(x_list)
                    ymax, ymin = max(y_list), min(y_list)
                
                write = file.write("\n")
            del x_list, y_list
            
        write = file.write("")    
        file.close()
        os.rename("xml/foruse.txt", "xml/" + data['imagePath'][:-4] + ".xml")

import os
import json
import cv2
import matplotlib.pyplot as plt

for i in os.listdir("json"):
    with open("json/" + i, encoding="utf-8") as f:
          
        data = json.load(f)

        #輸出縮排Json
        jsonData_sort = json.dumps(data, sort_keys = True, indent=4)  
    
        for j in range(0, len(data['shapes'])): 
            if (data['shapes'][j]['group_id'] == 1 or data['shapes'][j]['group_id'] == 4):#中文單字
                print(data['imagePath'])
                print(data['shapes'][j]['label'], data['shapes'][j]['group_id'])
                print(data['shapes'][j]['points'])
                a = data['shapes'][j]['label']
                a = str(a)
                
                print("---")
                
                x_list, y_list = [], []
                for points in range(0, 4):
                    x_list.append(data['shapes'][j]['points'][points][0])
                    y_list.append(data['shapes'][j]['points'][points][1])         
                xmax, xmin = max(x_list), min(x_list)
                ymax, ymin = max(y_list), min(y_list)
                
                if (ymin < 0):
                    ymin = 0
                if (xmin < 0):
                    xmin = 0
          
                img_ori = cv2.imread("img/" + data['imagePath'])   
                img_after = img_ori[ data['shapes'][j]['points'][0][1] : data['shapes'][j]['points'][2][1], data['shapes'][j]['points'][0][0] : data['shapes'][j]['points'][2][0]]

                cv2.imencode('.jpg', img_after)[1].tofile("part_img/" + a + "_" + data['imagePath'][:-4] + "_" + str(j) + ".jpg")
                
                del x_list, y_list

import os 
import shutil

cla = []
for i in os.listdir("part_img"):
    WordClass = i[0]
    if WordClass not in cla:
        cla.append(WordClass)
        os.mkdir("Class/" + WordClass)
        shutil.copy("part_img/" + i, "Class/" + WordClass + "/" + i)
    else:
        shutil.copy("part_img/" + i, "Class/" + WordClass + "/" + i)
        
#刪調資料小於九筆的類別
for j in os.listdir("Class/"):
    if (len(os.listdir("Class/" + j)) < 9):
        shutil.rmtree("Class/" + j)
    

import os
import cv2
import csv

print(os.listdir("AfterData"))

# 開啟 CSV 檔案
with open('Task2_Public_String_Coordinate.csv', newline='') as csvfile:

  # 讀取 CSV 檔案內容
  rows = csv.reader(csvfile)
  num = 0
  # 以迴圈輸出每一列
  for row in rows:
    num+=1
    x_list, y_list = [], []
    img_ori = cv2.imread("img_public/" + row[0] + ".jpg") 
    print(row[0])
    
    for i in range(1,len(row)):
        if (i%2==1):
            x_list.append(int(row[i]))
        else:
            y_list.append(int(row[i]))  

    xmax, xmin = max(x_list), min(x_list)
    ymax, ymin = max(y_list), min(y_list) 
    
    print(num, xmin, xmax, ymin, ymax)
    print(img_ori.shape)
    img_after = img_ori[int(ymin) : int(ymax), int(xmin) : int(xmax)]
    cv2.imwrite("AfterData/" + row[0] + "_" + str(num).zfill(4) + ".jpg", img_after)
    del x_list, y_list

2048 or img.shape[1] > 2048): back = cv2.resize(space, (3096, 3096)) back[1548 - round(y/2) : 1548 - round(y/2) + y, 1548 - round(x/2) : 1548 - round(x/2) + x] = img elif(img.shape[0] > 1024 or img.shape[1] > 1024): back = cv2.resize(space, (2048, 2048)) back[1024 - round(y/2) : 1024 - round(y/2) + y, 1024 - round(x/2) : 1024 - round(x/2) + x] = img elif(img.shape[0] <= 50 and img.shape[1] <= 50): img = cv2.resize(img, (x*4, y*4), interpolation=cv2.INTER_CUBIC) x = x*4 y = y*4 back = cv2.resize(space, (512, 512)) back[256 - round(y/2) : 256 - round(y/2) + y, 256 - round(x/2) : 256 - round(x/2) + x] = img elif(img.shape[0] <= 50 or img.shape[1] <= 50): le = x if (y > x): le = y if (le % 2 != 0): le = le + 101 else: le = le + 100 img = cv2.resize(img, (x*2, y*2), interpolation=cv2.INTER_CUBIC) x = x*2 y = y*2 back = cv2.resize(space, (le * 2, le * 2)) back[le - round(y/2) : le - round(y/2) + y, le - round(x/2) : le - round(x/2) + x] = img elif(img.shape[0] < 1025 and img.shape[1] < 1025): back = space back[512 - round(y/2) : 512 - round(y/2) + y, 512 - round(x/2) : 512 - round(x/2) + x] = img cv2.imwrite("target_data/" + i, back)">

import os
import numpy as np
import cv2

for i in os.listdir("part_data"):
    img = cv2.imread("part_data/" + i)
    space = cv2.imread("space1024.jpg")
    y = img.shape[0]
    x = img.shape[1]

    if(img.shape[0] > 2048 or img.shape[1] > 2048):
        
        back = cv2.resize(space, (3096, 3096))
        back[1548 - round(y/2) : 1548 - round(y/2) + y, 1548 - round(x/2) : 1548 - round(x/2) + x] = img
     
    elif(img.shape[0] > 1024 or img.shape[1] > 1024):
        
        back = cv2.resize(space, (2048, 2048))
        back[1024 - round(y/2) : 1024 - round(y/2) + y, 1024 - round(x/2) : 1024 - round(x/2) + x] = img
        
    elif(img.shape[0] <= 50 and img.shape[1] <= 50):
             
        img = cv2.resize(img, (x*4, y*4), interpolation=cv2.INTER_CUBIC)
        x = x*4
        y = y*4
        back = cv2.resize(space, (512, 512))
        back[256 - round(y/2) : 256 - round(y/2) + y, 256 - round(x/2) : 256 - round(x/2) + x] = img
    
    elif(img.shape[0] <= 50 or img.shape[1] <= 50):
        
        le = x
        if (y > x):
            le = y
        if (le % 2 != 0):
            le = le + 101
        else:
            le = le + 100
           
        img = cv2.resize(img, (x*2, y*2), interpolation=cv2.INTER_CUBIC)
        x = x*2
        y = y*2
        back = cv2.resize(space, (le * 2, le * 2))
        back[le - round(y/2) : le - round(y/2) + y, le - round(x/2) : le - round(x/2) + x] = img
     
    elif(img.shape[0] < 1025 and img.shape[1] < 1025):
        
        back = space
        back[512 - round(y/2) : 512 - round(y/2) + y, 512 - round(x/2) : 512 - round(x/2) + x] = img
    
    cv2.imwrite("target_data/" + i, back)

import os 
import shutil
import random

path = "img/"
all_data = os.listdir(path)
num = len(all_data)
os.mkdir("DetectTrain")
os.mkdir("DetectTrain/train")
os.mkdir("DetectTrain/validation")
os.mkdir("DetectTrain/train/images")
os.mkdir("DetectTrain/train/annotations")
os.mkdir("DetectTrain/validation/images")
os.mkdir("DetectTrain/validation/annotations")

percent = 0.85
select_list = random.sample(range(0, num), round(num * (1 - percent)))

for i in range(0, num):
    print(i + 1)
    if i in select_list:        
        shutil.copy("img/" + all_data[i], "DetectTrain/validation/images/" + all_data[i])
        shutil.copy("xml/" + all_data[i][:-4] + ".xml", "DetectTrain/validation/annotations/" + all_data[i][:-4] + ".xml")

    else:        
        shutil.copy("img/" + all_data[i], "DetectTrain/train/images/" + all_data[i])
        shutil.copy("xml/" + all_data[i][:-4] + ".xml", "DetectTrain/train/annotations/" + all_data[i][:-4] + ".xml")

import os
import shutil
import random

path = "Class/"
percent = 0.9 #train data
os.mkdir("ClassifyTrain")
os.mkdir("ClassifyTrain/train/")
os.mkdir("ClassifyTrain/valid/")
os.mkdir("ClassifyTrain/test/")

for i in os.listdir(path):
    num = len(os.listdir(path + i))
    select_list = random.sample(range(0, num), round(num * percent))
    select_list2 = random.sample(select_list, round(num * 0.9 * 0.8))
    fold = os.listdir(path + "/" + i)
    os.mkdir("ClassifyTrain/train/" + i)
    os.mkdir("ClassifyTrain/valid/" + i)
    os.mkdir("ClassifyTrain/test/" + i)
    
    for data in range(0, num):
        if data in select_list2:        
            shutil.copy("Class/" + i + "/" + fold[data], "ClassifyTrain/train/" + i + "/" + fold[data])
        elif data in select_list: 
            shutil.copy("Class/" + i + "/" + fold[data], "ClassifyTrain/valid/" + i + "/" + fold[data])
        else:  
            shutil.copy("Class/" + i + "/" + fold[data], "ClassifyTrain/test/" + i + "/" + fold[data])

from imageai.Detection.Custom import DetectionModelTrainer

trainer = DetectionModelTrainer()
trainer.setModelTypeAsYOLOv3()
trainer.setDataDirectory(data_directory = "DetectTrain") #指定到訓練資料夾
trainer.setTrainConfig(object_names_array = ["word", "NumLetter"],
                       batch_size = 4,
                       num_experiments = 60)
trainer.trainModel()

= 1: # 儲存訓練好的模型 路徑請自行調整 net_final.save(WEIGHTS_FINAL + ".h5")">

from tensorflow.keras import backend as K
from tensorflow.keras.models import Model
from tensorflow.keras.layers import Flatten, Dense, Dropout
from tensorflow.keras.applications.inception_resnet_v2 import InceptionResNetV2
from tensorflow.keras.optimizers import Adam
from tensorflow.keras.preprocessing.image import ImageDataGenerator

# 資料路徑
DATASET_PATH  = 'ClassifyTrain'

# 影像大小
IMAGE_SIZE = (100, 100)

# 影像類別數
NUM_CLASSES = 979

# 若 GPU 記憶體不足，可調降 batch size 或凍結更多層網路
BATCH_SIZE = 16

# 凍結網路層數
FREEZE_LAYERS = 2

# Epoch 數
NUM_EPOCHS = 2

# 模型輸出儲存的檔案
WEIGHTS_FINAL = 'model-resnet50-e150--No-zoom_range.h5'

# 透過 data augmentation 產生訓練與驗證用的影像資料
train_datagen = ImageDataGenerator(rotation_range=20,
                                   width_shift_range=0.2, #水平移
                                   height_shift_range=0.2, #垂直移
                                   shear_range=0.2, #x,y一個固定平移
                                   #zoom_range=0.2, #縮放
                                   channel_shift_range=10, #變色濾鏡
                                   #horizontal_flip=True, 垂直翻轉
                                   fill_mode='nearest')
train_batches = train_datagen.flow_from_directory(DATASET_PATH + '/train',
                                                  target_size=IMAGE_SIZE,
                                                  interpolation='bicubic',
                                                  class_mode='categorical',
                                                  shuffle=True,
                                                  batch_size=BATCH_SIZE)

valid_datagen = ImageDataGenerator()
valid_batches = valid_datagen.flow_from_directory(DATASET_PATH + '/valid',
                                                  target_size=IMAGE_SIZE,
                                                  interpolation='bicubic',
                                                  class_mode='categorical',
                                                  shuffle=False,
                                                  batch_size=BATCH_SIZE)

# 輸出各類別的索引值
for cls, idx in train_batches.class_indices.items():
    print('Class #{} = {}'.format(idx, cls))

# 以訓練好的 InceptionResNetV2 為基礎來建立模型，
# 捨棄 InceptionResNetV2 頂層的 fully connected layers
net = InceptionResNetV2(include_top=False, weights='imagenet', input_tensor=None,
               input_shape=(IMAGE_SIZE[0],IMAGE_SIZE[1],3))
x = net.output
x = Flatten()(x)

# 增加 DropOut layer
x = Dropout(0.5)(x)

# 增加 Dense layer，以 softmax 產生個類別的機率值
output_layer = Dense(NUM_CLASSES, activation='softmax', name='softmax')(x)

# 設定凍結與要進行訓練的網路層
net_final = Model(inputs=net.input, outputs=output_layer)
for layer in net_final.layers[:FREEZE_LAYERS]:
    layer.trainable = False
for layer in net_final.layers[FREEZE_LAYERS:]:
    layer.trainable = True

# 使用 Adam optimizer，以較低的 learning rate 進行 fine-tuning
net_final.compile(optimizer=Adam(lr=1e-5),
                  loss='categorical_crossentropy', metrics=['accuracy'])

# 訓練模型
for i in range(1,101):
    WEIGHTS_FINAL = "model-InceptionResnetV2-e" + str(i * 2)
    net_final.fit_generator(train_batches,
                        steps_per_epoch = train_batches.samples // BATCH_SIZE,
                        validation_data = valid_batches,
                        validation_steps = valid_batches.samples // BATCH_SIZE,
                        epochs = NUM_EPOCHS)
    if i >= 1:
        # 儲存訓練好的模型 路徑請自行調整
        net_final.save(WEIGHTS_FINAL + ".h5")

0): tar_list = [] for detection in detections: if (detection["name"]=="word"): tar_list.append(detection["box_points"]) # xmin ymin xmax ymax length = len(tar_list) if (length > 0): tar_list.sort(key = lambda x:x[1])#Y排序 Yrange = tar_list[length-1][3] - tar_list[0][1] tar_list.sort(key = lambda x:x[0])#X排序 Xrange = tar_list[length-1][2] - tar_list[0][0] if (Xrange > Yrange): for n in tar_list: out = img[n[1]:n[3], n[0]:n[2]] o = test(cls_list, net, out) ans = ans + o else: tar_list.sort(key = lambda x:x[1]) for n in tar_list: out = img[n[1]:n[3], n[0]:n[2]] o = test(cls_list, net, out) ans = ans + o with open('output.csv', 'a', newline='', encoding="utf-8") as csvfile: writer = csv.writer(csvfile) writer.writerow(ans) csvfile.close() else: with open('output.csv', 'a', newline='', encoding="utf-8") as csvfile: writer = csv.writer(csvfile) writer.writerow(["###"]) csvfile.close() else: with open('output.csv', 'a', newline='', encoding="utf-8") as csvfile: writer = csv.writer(csvfile) writer.writerow(["###"]) csvfile.close()">

from imageai.Detection.Custom import CustomObjectDetection
from tensorflow.python.keras import backend as K
from tensorflow.python.keras.models import load_model
from tensorflow.python.keras.preprocessing import image
from tensorflow.keras.layers import Flatten, Dense, Dropout
from tensorflow.keras.models import Model
import tensorflow as tf
import sys
import numpy as np
import os
import shutil
import random
import cv2
import csv

#要改成你所選的模型路徑名稱
ImageAIJsonPath = "DetectTrain/json/detection_config.json"
ImageAIModelPath = "DetectTrain/models/xxx.h5"
InceptionResNetV2Path = 'model-InceptionResNetV2-e102.h5'



def test(cls_list, net, img):

        img = cv2.resize(img, (100, 100), interpolation=cv2.INTER_CUBIC)
        if img is None:
            return 0
        x = image.img_to_array(img)
        x = np.expand_dims(x, axis = 0)
        pred = net_final.predict(x)[0]
        top_inds = pred.argsort()[::-1][:3]

        return cls_list[top_inds[0]]
        
cls_list = os.listdir("ClassifyTrain/test")
# 影像大小
IMAGE_SIZE = (100, 100)

# 影像類別數
NUM_CLASSES = 979

# 凍結網路層數
FREEZE_LAYERS = 2

start = 3

#類別載入
cls_list = cls_list
#模組載入
for i in range(0,1):
    net = tf.keras.applications.InceptionResNetV2(include_top=False, weights='imagenet', input_tensor=None,
               input_shape=(IMAGE_SIZE[0], IMAGE_SIZE[1], 3))
    x = net.output
    x = Flatten()(x)

    # 增加 DropOut layer
    x = Dropout(0.5)(x)
    
    # 增加 Dense layer，以 softmax 產生個類別的機率值
    output_layer = Dense(NUM_CLASSES , activation='softmax', name='softmax')(x)

    # 設定凍結與要進行訓練的網路層
    net_final = Model(inputs=net.input, outputs=output_layer)
    for layer in net_final.layers[:FREEZE_LAYERS]:
        layer.trainable = False
    for layer in net_final.layers[FREEZE_LAYERS:]:
        layer.trainable = True
    
    # 載入舊權重 記得更改到所選的路徑與模型名稱
    net_final.load_weights(InceptionResNetV2Path) 


detector = CustomObjectDetection()
detector.setModelTypeAsYOLOv3()
detector.setModelPath(ImageAIModelPath) 
detector.setJsonPath(ImageAIJsonPath)
detector.loadModel()

for i in os.listdir("target_data/"): 
    
    img = cv2.imread("target_data/" + i)
    detections = detector.detectObjectsFromImage(input_image = "target_data/" + i,
                                             output_image_path = "output.jpg",
                                             minimum_percentage_probability = 50)
    ans = ""
    if (len(detections) > 0):
        
        tar_list = []
        for detection in detections:
            if (detection["name"]=="word"):
                tar_list.append(detection["box_points"]) # xmin ymin xmax ymax
        
        length = len(tar_list)
        if (length > 0):      
        
            tar_list.sort(key = lambda x:x[1])#Y排序
            Yrange = tar_list[length-1][3] - tar_list[0][1]
            tar_list.sort(key = lambda x:x[0])#X排序
            Xrange = tar_list[length-1][2] - tar_list[0][0]
        
            if (Xrange > Yrange):
            
                for n in tar_list:
                    out = img[n[1]:n[3], n[0]:n[2]]

                    o = test(cls_list, net, out)
                    ans = ans + o
            else:
                tar_list.sort(key = lambda x:x[1])
                for n in tar_list:
                    out = img[n[1]:n[3], n[0]:n[2]]

                    o = test(cls_list, net, out)
                    ans = ans + o
            
            with open('output.csv', 'a', newline='', encoding="utf-8") as csvfile:
                writer = csv.writer(csvfile)
                writer.writerow(ans)

            csvfile.close()
        else:
            with open('output.csv', 'a', newline='', encoding="utf-8") as csvfile:
                writer = csv.writer(csvfile)
                writer.writerow(["###"])

            csvfile.close()
    else:
        with open('output.csv', 'a', newline='', encoding="utf-8") as csvfile:
            writer = csv.writer(csvfile)
            writer.writerow(["###"])

        csvfile.close()