导读
本文主要介绍使用Python-OpenCV实现餐盘水果识别与计价的应用。
测试图像与说明
使用图像如下，拍摄环境有待改善(存在光照不均和拍摄角度的影响):
餐盘/菜品识别一般方法：
(1)识别餐盘---传统方法和机器学习/深度学习方法；
(2)识别菜品---机器学习/深度学习方法；
本文使用传统方法识别餐盘。
效果演示：
算法思路与实现步骤
思路：传统方法识别餐盘---依据颜色和形状来区分。
具体步骤：
(1)餐盘颜色共三种：白色、绿色、橙色，形状共两种：圆形和方形。区别颜色使用HSV阈值范围筛选即可，圆形与方形通过轮廓面积与轮廓最小外接圆面积的比值来筛选，圆形rate>=0.9，方形<0.9;
(2)水果共三种：苹果、香蕉、橙子，通过颜色可以区分苹果和橙子，通过轮廓最小外接矩形的宽高比可以区分香蕉和橙子；
(3)计价：盘子和水果的数量乘以对应的单价即可；
(4)设计UI，计价时显示收款码。
Python-OpenCV实现算法核心代码与效果如下：
def Recognize_Dish(self,img):
#-------------------香蕉检测-----------------#
banana_num = 0
hsv_img=cv2.cvtColor(img,cv2.COLOR_BGR2HSV)
lower_yellow = np.array([15,30,145])#颜色范围低阈值
upper_yellow = np.array([35,255,255])#颜色范围高阈值
mask = cv2.inRange(hsv_img,lower_yellow,upper_yellow)#根据颜色范围删选
mask = cv2.medianBlur(mask, 5)#中值滤波
#cv2.imshow('mask_banana', mask)
contours,hierarchy = cv2.findContours(mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_NONE)
for cnt in contours:
rect = cv2.minAreaRect(cnt)
box = cv2.boxPoints(rect)
box = np.int0(box)
width = max(rect[1][0],rect[1][1])
height = min(rect[1][0],rect[1][1])
center = (int(rect[0][0]),int(rect[0][1]))
if width > 180 and height > 80 and height < 130:
#print(width,height)
img = cv2.drawContours(img,[box],0,(0,0,255),2)
cv2.putText(img,'banana',center,font,1,(255,0,255), 2)
banana_num += 1
item_0 = QTableWidgetItem("%d"%banana_num)
self.tableWidget.setItem(8, 0, item_0)
#-------------------苹果检测-----------------#
apple_num = 0
lower_apple = np.array([0,50,50])#颜色范围低阈值
upper_apple = np.array([30,255,255])#颜色范围高阈值
mask_apple = cv2.inRange(hsv_img,lower_apple,upper_apple)#根据颜色范围删选
mask_apple = cv2.medianBlur(mask_apple, 9)#中值滤波
#cv2.imshow('mask_apple', mask_apple)
#cv2.imwrite('mask_apple.jpg', mask_apple)
contours2,hierarchy2 = cv2.findContours(mask_apple, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_NONE)
for cnt2 in contours2:
center,radius = cv2.minEnclosingCircle(cnt2)
area = cv2.contourArea(cnt2)
#print(radius)
rate = area / (math.pi * radius *radius)
if radius > 50 and radius < 75 and rate < 0.91:
#print(radius)
cv2.circle(img,(int(center[0]),int(center[1])),int(radius),(0,255,0),2)
cv2.putText(img,'apple',(int(center[0]),int(center[1])),font,1,(255,0,0), 2)
apple_num += 1
item_1 = QTableWidgetItem("%d"%apple_num)
self.tableWidget.setItem(6, 0, item_1)
#-------------------橘子检测-----------------#
orange_num = 0
lower_orange = np.array([0,90,60])#颜色范围低阈值
upper_orange = np.array([60,255,255])#颜色范围高阈值
mask_orange = cv2.inRange(hsv_img,lower_orange,upper_orange)#根据颜色范围删选
mask_orange = cv2.medianBlur(mask_orange, 5)#中值滤波
#cv2.imshow('mask_orange', mask_orange)
#cv2.imwrite('mask_orange.jpg', mask_orange)
contours3,hierarchy3 = cv2.findContours(mask_orange, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_NONE)
for cnt3 in contours3:
center,radius = cv2.minEnclosingCircle(cnt3)
area = cv2.contourArea(cnt3)
#print(radius)
rate = area / (math.pi * radius *radius)
if radius > 50 and radius < 75 and rate > 0.85:
#print(radius)
cv2.circle(img,(int(center[0]),int(center[1])),int(radius),(255,0,255),2)
cv2.putText(img,'orange',(int(center[0]),int(center[1])),font,1,(255,255,0), 2)
orange_num += 1
item_2 = QTableWidgetItem("%d"%orange_num)
self.tableWidget.setItem(7, 0, item_2)
#-------------------白色餐盘检测-----------------#
white_circle_num = 0
white_rect_num = 0
lower_white = np.array([0,0,150])#颜色范围低阈值
upper_white= np.array([100,55,255])#颜色范围高阈值
mask_white = cv2.inRange(hsv_img,lower_white,upper_white)#根据颜色范围删选
mask_white = cv2.medianBlur(mask_white, 5)#中值滤波
#cv2.imshow('mask_white', mask_white)
#cv2.imwrite('mask_white.jpg', mask_white)
contours4,hierarchy4 = cv2.findContours(mask_white, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_NONE)
for cnt4 in contours4:
area = cv2.contourArea(cnt4)
center,radius = cv2.minEnclosingCircle(cnt4)
#print(radius)
rate = area / (math.pi * radius *radius)
if radius > 100 and radius < 160:
#print(radius)
if rate >= 0.9:
cv2.circle(img,(int(center[0]),int(center[1])),int(radius),(255,255,0),2)
cv2.putText(img,'white_circle',(int(center[0]),int(center[1])),font,1,(0,255,0), 2)
white_circle_num += 1
elif rate >0.6 and rate < 0.9:
rect = cv2.minAreaRect(cnt4)
box = cv2.boxPoints(rect)
box = np.int0(box)
#cv2.circle(img,(int(center[0]),int(center[1])),int(radius),(255,0,255),5)
img = cv2.drawContours(img,[box],0,(255,255,0),2)
cv2.putText(img,'white_rect',(int(center[0]),int(center[1])),font,1,(0,255,0), 2)
white_rect_num += 1
item_3 = QTableWidgetItem("%d"%white_circle_num)
self.tableWidget.setItem(0, 0, item_3)
item_4 = QTableWidgetItem("%d"%white_rect_num)
self.tableWidget.setItem(1, 0, item_4)
#-------------------绿色餐盘检测-----------------#
green_circle_num = 0
green_rect_num = 0
lower_green = np.array([30,65,65])#颜色范围低阈值
upper_green= np.array([80,255,255])#颜色范围高阈值
mask_green = cv2.inRange(hsv_img,lower_green,upper_green)#根据颜色范围删选
mask_green = cv2.medianBlur(mask_green, 5)#中值滤波
#cv2.imshow('mask_green', mask_green)
#cv2.imwrite('mask_green.jpg', mask_green)
contours5,hierarchy5 = cv2.findContours(mask_green, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_NONE)
for cnt5 in contours5:
area = cv2.contourArea(cnt5)
center,radius = cv2.minEnclosingCircle(cnt5)
#print(radius)
rate = area / (math.pi * radius *radius)
if radius > 100 and radius < 160:
#print(radius)
if rate >= 0.9:
cv2.circle(img,(int(center[0]),int(center[1])),int(radius),(0,255,0),2)
cv2.putText(img,'green_circle',(int(center[0]),int(center[1])),font,1,(0,255,255), 2)
green_circle_num += 1
elif rate >0.6 and rate < 0.9:
rect = cv2.minAreaRect(cnt5)
box = cv2.boxPoints(rect)
box = np.int0(box)
#cv2.circle(img,(int(center[0]),int(center[1])),int(radius),(255,0,255),5)
img = cv2.drawContours(img,[box],0,(0,255,0),2)
cv2.putText(img,'green_rect',(int(center[0]),int(center[1])),font,1,(0,255,255), 2)
green_rect_num += 1
item_5 = QTableWidgetItem("%d"%green_circle_num)
self.tableWidget.setItem(4, 0, item_5)
item_6 = QTableWidgetItem("%d"%green_rect_num)
self.tableWidget.setItem(5, 0, item_6)
#-------------------橙色餐盘检测-----------------#
orange_circle_num = 0
orange_rect_num = 0
lower_orange_dish = np.array([0,100,100])#颜色范围低阈值
upper_orange_dish= np.array([15,255,255])#颜色范围高阈值
mask_orange_dish = cv2.inRange(hsv_img,lower_orange_dish,upper_orange_dish)#根据颜色范围删选
mask_orange_dish = cv2.medianBlur(mask_orange_dish, 5)#中值滤波
#cv2.imshow('mask_green', mask_green)
#cv2.imwrite('mask_orange_dish.jpg', mask_orange_dish)
contours6,hierarchy6 = cv2.findContours(mask_orange_dish, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_NONE)
for cnt6 in contours6:
area = cv2.contourArea(cnt6)
center,radius = cv2.minEnclosingCircle(cnt6)
#print('----------------')
#print(radius)
rate = area / (math.pi * radius *radius)
if radius > 100 and radius < 160:
#print(rate)
if rate >= 0.8:
cv2.circle(img,(int(center[0]),int(center[1])),int(radius),(0,255,0),2)
cv2.putText(img,'orange_circle',(int(center[0]),int(center[1])),font,1,(255,0,255), 2)
orange_circle_num += 1
elif rate >0.3 and rate < 0.8:
rect = cv2.minAreaRect(cnt6)
box = cv2.boxPoints(rect)
box = np.int0(box)
#cv2.circle(img,(int(center[0]),int(center[1])),int(radius),(255,0,255),5)
img = cv2.drawContours(img,[box],0,(0,255,0),2)
cv2.putText(img,'orange_rect',(int(center[0]),int(center[1])),font,1,(255,0,255), 2)
orange_rect_num += 1
item_7 = QTableWidgetItem("%d"%orange_circle_num)
self.tableWidget.setItem(2, 0, item_7)
item_8 = QTableWidgetItem("%d"%orange_rect_num)
self.tableWidget.setItem(3, 0, item_8)
for i in range(0,9):
self.tableWidget.item(i,0).setTextAlignment(QtCore.Qt.AlignHCenter|QtCore.Qt.AlignVCenter)
self.tableWidget.item(i,1).setTextAlignment(QtCore.Qt.AlignHCenter|QtCore.Qt.AlignVCenter)
#----------------计算价格--------------#
self.price = self.price_white_circle * white_circle_num + \
self.price_white_rect * white_rect_num + \
self.price_orange_circle * orange_circle_num + \
self.price_orange_rect * orange_rect_num + \
self.price_green_circle * green_circle_num + \
self.price_green_rect * green_rect_num + \
self.price_apple * apple_num + \
self.price_orange * orange_num +\
self.price_banana * banana_num
print(self.price)
return img
结尾语
(1) 算法只针对水果和餐盘数量和形态较少的情形，方法供参考；
(2) 实际应用将更复杂，要求更高，一般开源的目标检测网络也很难满足要求；
(3) 常见菜品识别的实际应用要求：一个菜只用一张图片训练或做模板，训练和识别时间尽量短，能够及时更新使用。所以真正类似的产品并不好做，如果你有好的方法欢迎留言。
点击下方链接关注我们，更多视觉/图像处理相关内容及时掌握。
