TensorFlow2.X结合OpenCV 实现手势识别功能

使用Tensorflow 构建卷积神经网络，训练手势识别模型，使用opencv DNN 模块加载模型实时手势识别
效果如下：

先显示下部分数据集图片（0到9的表示，感觉很怪）

构建模型进行训练

数据集地址

import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import datasets,layers,optimizers,Sequential,metrics
from tensorflow.python.framework.convert_to_constants import convert_variables_to_constants_v2
import os 
import pathlib
import random
import matplotlib.pyplot as plt
os.environ['TF_CPP_MIN_LOG_LEVEL']='2'
def read_data(path):
 path_root = pathlib.Path(path)
 # print(path_root)
 # for item in path_root.iterdir():
 #  print(item)
 image_paths = list(path_root.glob('*/*'))
 image_paths = [str(path) for path in image_paths]
 random.shuffle(image_paths)
 image_count = len(image_paths)
 # print(image_count)
 # print(image_paths[:10])
 label_names = sorted(item.name for item in path_root.glob('*/') if item.is_dir())
 # print(label_names)
 label_name_index = dict((name, index) for index, name in enumerate(label_names))
 # print(label_name_index)
 image_labels = [label_name_index[pathlib.Path(path).parent.name] for path in image_paths]
 # print("First 10 labels indices: ", image_labels[:10])
 return image_paths,image_labels,image_count
def preprocess_image(image):
 image = tf.image.decode_jpeg(image, channels=3)
 image = tf.image.resize(image, [100, 100])
 image /= 255.0 # normalize to [0,1] range
 # image = tf.reshape(image,[100*100*3])
 return image
def load_and_preprocess_image(path,label):
 image = tf.io.read_file(path)
 return preprocess_image(image),label
def creat_dataset(image_paths,image_labels,bitch_size):
 db = tf.data.Dataset.from_tensor_slices((image_paths, image_labels))
 dataset = db.map(load_and_preprocess_image).batch(bitch_size) 
 return dataset
def train_model(train_data,test_data):
 #构建模型
 network = keras.Sequential([
   keras.layers.Conv2D(32,kernel_size=[5,5],padding="same",activation=tf.nn.relu),
   keras.layers.MaxPool2D(pool_size=[2, 2], strides=2, padding='same'),
   keras.layers.Conv2D(64,kernel_size=[3,3],padding="same",activation=tf.nn.relu),
   keras.layers.MaxPool2D(pool_size=[2, 2], strides=2, padding='same'),
   keras.layers.Conv2D(64,kernel_size=[3,3],padding="same",activation=tf.nn.relu),
   keras.layers.Flatten(),
   keras.layers.Dense(512,activation='relu'),
   keras.layers.Dropout(0.5),
   keras.layers.Dense(128,activation='relu'),
   keras.layers.Dense(10)])
 network.build(input_shape=(None,100,100,3))
 network.summary()
 network.compile(optimizer=optimizers.SGD(lr=0.001),
   loss=tf.losses.SparseCategoricalCrossentropy(from_logits=True),
   metrics=['accuracy']
 )
 #模型训练
 network.fit(train_data, epochs = 100,validation_data=test_data,validation_freq=2) 
 network.evaluate(test_data)
 tf.saved_model.save(network,'D:\\code\\PYTHON\\gesture_recognition\\model\\')
 print("保存模型成功")
 # Convert Keras model to ConcreteFunction
 full_model = tf.function(lambda x: network(x))
 full_model = full_model.get_concrete_function(
 tf.TensorSpec(network.inputs[0].shape, network.inputs[0].dtype))
 # Get frozen ConcreteFunction
 frozen_func = convert_variables_to_constants_v2(full_model)
 frozen_func.graph.as_graph_def()

 layers = [op.name for op in frozen_func.graph.get_operations()]
 print("-" * 50)
 print("Frozen model layers: ")
 for layer in layers:
  print(layer)

 print("-" * 50)
 print("Frozen model inputs: ")
 print(frozen_func.inputs)
 print("Frozen model outputs: ")
 print(frozen_func.outputs)

 # Save frozen graph from frozen ConcreteFunction to hard drive
 tf.io.write_graph(graph_or_graph_def=frozen_func.graph,
   logdir="D:\\code\\PYTHON\\gesture_recognition\\model\\frozen_model\\",
   name="frozen_graph.pb",
   as_text=False)
 print("模型转换完成，训练结束")


if __name__ == "__main__":
 print(tf.__version__)
 train_path = 'D:\\code\\PYTHON\\gesture_recognition\\Dataset'
 test_path = 'D:\\code\\PYTHON\\gesture_recognition\\testdata' 
 image_paths,image_labels,_ = read_data(train_path)
 train_data = creat_dataset(image_paths,image_labels,16)
 image_paths,image_labels,_ = read_data(test_path)
 test_data = creat_dataset(image_paths,image_labels,16)
 train_model(train_data,test_data)

OpenCV加载模型，实时检测

这里为了简化检测使用了ROI。

import cv2
from cv2 import dnn
import numpy as np
print(cv2.__version__)
class_name = ['0', '1', '2', '3', '4', '5', '6', '7', '8', '9']
net = dnn.readNetFromTensorflow('D:\\code\\PYTHON\\gesture_recognition\\model\\frozen_model\\frozen_graph.pb')
cap = cv2.VideoCapture(0)
i = 0
while True:
 _,frame= cap.read() 
 src_image = frame
 cv2.rectangle(src_image, (300, 100),(600, 400), (0, 255, 0), 1, 4)
 frame = cv2.cvtColor(frame,cv2.COLOR_BGR2RGB)
 pic = frame[100:400,300:600]
 cv2.imshow("pic1", pic)
 # print(pic.shape)
 pic = cv2.resize(pic,(100,100))
 blob = cv2.dnn.blobFromImage(pic,  
        scalefactor=1.0/225.,
        size=(100, 100),
        mean=(0, 0, 0),
        swapRB=False,
        crop=False)
 # blob = np.transpose(blob, (0,2,3,1))       
 net.setInput(blob)
 out = net.forward()
 out = out.flatten()

 classId = np.argmax(out)
 # print("classId",classId)
 print("预测结果为：",class_name[classId])
 src_image = cv2.putText(src_image,str(classId),(300,100), cv2.FONT_HERSHEY_SIMPLEX, 2,(0,0,255),2,4)
 # cv.putText(img, text, org, fontFace, fontScale, fontcolor, thickness, lineType)
 cv2.imshow("pic",src_image)
 if cv2.waitKey(10) == ord('0'):
  break

小结

这里本质上还是一个图像分类任务。而且，样本数量较少。优化的时候需要做数据增强，还需要防止过拟合。