基于CNN的静态手势识别系统

# -*- coding: utf-8 -*- from PyQt4.QtGui import * from PyQt4.QtCore import * from PyQt4.QtSql import * import os import sys import copy from PIL import Image import time import cv2 import qimage2ndarray import mmap import numpy as np import multiprocessing import re import random sample_counts = 400 skinkernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (5, 5)) # 工具类（单例模式） class UtilOperation(object): _shared_state = {} mGesture = None mm = None def __new__(cls, *args, **kwargs): obj = super(UtilOperation, cls).__new__(cls, *args, **kwargs) obj.__dict__ = cls._shared_state print('__new__') return obj def gestureLoadNNByBtn(self, weight_index=0, weight_name='binmask_CnnModel.hdf5'): # xunfengd # from train_gesture_model import TrainGestureModeByCNN from gesture_recognize import GestureRecognize from keras import backend as K K.clear_session() # clear session # xunfengd get gesture object self.mGesture = GestureRecognize(-1) self.mGesture.loadCNN(weight_index, weight_name) self.mGesture.static_image_recognize('train_set/eight/eight_24.png') print('gestureLoadNN...' + weight_name) return self.mGesture def gestureLoadNN(self, weight_index=0, weight_name='binmask_CnnModel.hdf5'): if self.mGesture is None: from gesture_recognize import GestureRecognize from keras import backend as K K.clear_session() # clear session self.mGesture = GestureRecognize() self.mGesture.loadCNN(weight_index, weight_name) # 手动预测一次，避免keras报错 self.mGesture.static_image_recognize('train_set/eight/eight_24.png') print('gestureLoadNN...' + weight_name) return self.mGesture def getMmap(self): if self.mm == None: self.mm = mmap.mmap(fileno=-1, length=1024 * 1024, access=mmap.ACCESS_READ, tagname='share_mmap') return self.mm # 训练CNN网络线程 class TrainCNNThread(QThread): sendlog = pyqtSignal(object) def __init__(self, mUtil, parent=None): super(TrainCNNThread, self).__init__(parent) self.stoped = False self.mUtil = mUtil self.mutex = QMutex() def close_keras_sesion(self): print('close_keras_sesion...') from keras import backend as K K.clear_session() self.exit(0) pass def run(self): with QMutexLocker(self.mutex): print('in run method,start to send signal!') self.stoped = False from train_gesture_model import TrainGestureModeByCNN train_net_obj = TrainGestureModeByCNN() train_net_obj.trainModel(train_set_path='train_set', weight_name='myNewCnnModel.hdf5') self.sendlog.emit(' ') # 静态图像采集线程 class StaticVideoCollectionThread(QThread): sendlog = pyqtSignal(object) sendlog_recog = pyqtSignal(object) # 使用二值化模型 def binaryMask(self, frame, x0, y0, width, height): # print('use binaryMask model ...') minValue = 70 # 创建矩形框 cv2.rectangle(frame, (x0, y0), (x0 + width, y0 + height), (0, 255, 0), 1) roi = frame[y0:y0 + height, x0:x0 + width] # 获取灰度图像 gray = cv2.cvtColor(roi, cv2.COLOR_BGR2GRAY) # 高斯模糊:高斯滤波器中像素的权重与其距中心像素的距离成比例 blur = cv2.GaussianBlur(gray, (5, 5), 2) # 图像的二值化提取目标,动态自适应的调整属于自己像素点的阈值，而不是整幅图像都用一个阈值 th3 = cv2.adaptiveThreshold(blur, 255, cv2.ADAPTIVE_THRESH_GAUSSIAN_C, cv2.THRESH_BINARY_INV, 11, 2) ret, res = cv2.threshold(th3, minValue, 255, cv2.THRESH_BINARY_INV + cv2.THRESH_OTSU) return res # 使用肤色检测模型 def skinMask(self, frame, x0, y0, width, height): print('use skin model ...') # HSV values low_range = np.array([0, 50, 80]) upper_range = np.array([30, 200, 255]) cv2.rectangle(frame, (x0, y0), (x0 + width, y0 + height), (0, 255, 0), 1) roi = frame[y0:y0 + height, x0:x0 + width] hsv = cv2.cvtColor(roi, cv2.COLOR_BGR2HSV) # 设阈值，去除背景部分,低于这个lower_red的值，图像值变为0,高于这个upper_red的值，图像值变为0 mask = cv2.inRange(hsv, low_range, upper_range) #腐蚀操作，减少整幅图像的白色区域 mask = cv2.erode(mask, skinkernel, iterations=1) #膨胀操作，增加图像中的白色区域 mask = cv2.dilate(mask, skinkernel, iterations=1) # 用高斯分布权值矩阵与原始图像矩阵做卷积运算 mask = cv2.GaussianBlur(mask, (15, 15), 1) # cv2.imshow("Blur", mask) # 图像与运算 res = cv2.bitwise_and(roi, roi, mask=mask) # color to grayscale res = cv2.cvtColor(res, cv2.COLOR_BGR2GRAY) return res def __init__(self, video_recongize_obj, parent=None): super(StaticVideoCollectionThread, self).__init__(parent) self.stoped = False self.mutex = QMutex() self.cap = None self.mm = mmap.mmap(fileno=-1, length=1024 * 1024, access=mmap.ACCESS_WRITE, tagname='share_mmap') self.video_recongize_obj = video_recongize_obj def run(self): with QMutexLocker(self.mutex): print('in run method,start to send signal!') self.stoped = False # if self.cap == None : self.cap = cv2.VideoCapture(0) ret = self.cap.set(3, 640) ret = self.cap.set(4, 480) while True: if self.stoped: break ret, frame = self.cap.read() max_area = 0 frame = cv2.flip(frame, 3) x0 = 400 y0 = 200 height = 200 width = 200 # 使用肤色检测模型 if self.video_recongize_obj.flag_enable_skin_extraction_model: roi_frame = self.skinMask(frame, x0, y0, width, height) else: # 使用高斯二值化模型 roi_frame = self.binaryMask(frame, x0, y0, width, height) fram_dict = {} fram_dict['frame'] = frame fram_dict['roi_frame'] = roi_frame self.mm.seek(0) frame_size = len(frame.tostring()) roi_frame_size = len(roi_frame.tostring()) self.mm.write(frame.tostring()) self.mm.write(roi_frame.tostring()) self.sendlog.emit(str(frame_size)) self.sendlog_recog.emit(str(roi_frame_size)) # print('start to send signal!') time.sleep(0.3) self.cap.release() def stop(self): with QMutexLocker(self.mutex): self.stoped = True def isStoped(self): with QMutexLocker(self.mutex): return self.stoped # 批量处理静态图片识别线程 class BatchProcessImgsThread(QThread): sendlog = pyqtSignal(object) def __init__(self, pic_recog_obj, mUtil, parent=None): super(BatchProcessImgsThread, self).__init__(parent) self.stoped = False self.mutex = QMutex() self.pic_recog_obj = pic_recog_obj self.mUtil = mUtil def run(self): with QMutexLocker(self.mutex): print('in BatchProcessImgsThread run method,start to send signal!') self.stoped = False if self.mUtil.mGesture is None: self.mUtil.gestureLoadNN() self.pic_recog_obj.getImgListPath(set_path=self.pic