Comparison of color detection methods in OpenCV inRange, numpy, cupy

Introduction

When playing with OpenCV, you may want to detect objects based on color. Often when doing color detection with OpenCV

--Convert from RGB color space to HSV color space using cv2.cvtColor --Specify the HSV color space range with cv2.inRange and binarize it. --Find Contours from the image that came out and filter by shape

The method is introduced. On the other hand, similarly, it is possible to use Numpy to binarize the conditions for each pixel. Here, we compared the speeds on Xavier NX, including the advantages and disadvantages of these two methods and the implementation in cupy.

What i did

We compared the following 4 conditions including the method of binarizing inRange after converting to HSV.

1. How to detect a green ball using inRange
2. How to binarize by specifying conditions with numpy 3, how to detect under conditions that can not be done with inRange with numpy When speeding up 4 and 3 with CUPY I compared the resulting images with 1 and 23 and the speeds with 1234, respectively. The source I tried with python3 is attached in 4 ways in the latter half.

Source image

Problems with inRange

Normally, when performing color detection using inRange, we want to perform color detection based on hue, so I think that we should first convert from RGB color space to HSV color space. in this case

--InRange basically has a low degree of freedom because it can only take a threshold value with a constant parallel to each axis in the color space. --When using inRange in HSV color space, the color around red straddles the boundary between H = 0 and H = 180, so it cannot be filtered with a single inRange. --H fluctuates around white and black depending on the lighting conditions. , --InRange cannot use CUDA acceleration

As a result, when using inRange As you can see, it will be very difficult to distinguish the black cable at the bottom of the image and the white wall at the top of the image from the green ball.

Implementation in Numpy

Filters similar to inRange in Numpy

numpy1.py

        hsv = cv2.cvtColor( frame , cv2.COLOR_BGR2HSV )

        h = frame[:,:,0]
        s = frame[:,:,1]
        v = frame[:,:,2]

        mask_g = np.zeros(h.shape, dtype=np.uint8)
        mask_g[ (h>20) & (h <100) & (s>200) & (s < 255) & (v>50) & (v<150 ) ] = 255

I think it can be implemented as. in this case

        mask_g[ (h>20) & (h <100) & (s>200) & (s < 255) & (v>50) & (v<150 ) ] = 255

Part is

        mask_g[ ( g/r> 2.0) & (g/b>2.0) ] = 255

Not only the threshold value as a constant parallel to the axis of the color space, but also the threshold value can be set based on the ratio of each element, and the linear or more complicated threshold value can be set. In the above example, the image after detection will be as follows, and I think that the expected result can be easily obtained. In the above example, it is detected when the green element is twice as bright as red and blue, and the black and white parts can be easily excluded as shown below.

Speed comparison

On the other hand, regarding speed,

1. inRange
2. numpy (filter condition similar to inRange)
3. numpy( R/B > 2 & R/G>2 )
4. cupy( R/B >2 & R/G > 2 ) It became as follows under the four conditions of.

I didn't think inRange was the fastest. .. .. ..

Source code I tried

1, in Range detects green ball

inrange.py

import cv2
import numpy as np
import time

src = 'v4l2src device= /dev/video0 ! image/jpeg,width=1920,height=1080 !jpegdec !videoconvert ! appsink'

cap=cv2.VideoCapture(src)


W = cap.get(cv2.CAP_PROP_FRAME_WIDTH)
H = cap.get(cv2.CAP_PROP_FRAME_HEIGHT)
fps = cap.get(cv2.CAP_PROP_FPS)
while(cap.isOpened()):
    sum=0
    for i in range( 0,100 ):
        ret, frame = cap.read()
        start=time.time()
        r = frame[:,:,0]
        g = frame[:,:,1]
        b = frame[:,:,2]

        mask_g = np.zeros(r.shape, dtype=np.uint8)
        mask_g[ ( g/r> 2.0) & (g/b>2.0) ] = 255
        end = time.time()
        sum+= (end- start)
        cv2.imshow('ReadM',mask_g )
        cv2.waitKey(1)
        if i % 10 == 0 :
            print( i )
    print( sum/100)
cap.release()

2. How to binarize by specifying conditions with numpy

numpy1.py

import cv2
import numpy as np
import time

src = 'v4l2src device= /dev/video0 ! image/jpeg,width=1920,height=1080 !jpegdec !videoconvert ! appsink'

cap=cv2.VideoCapture(src)


W = cap.get(cv2.CAP_PROP_FRAME_WIDTH)
H = cap.get(cv2.CAP_PROP_FRAME_HEIGHT)
fps = cap.get(cv2.CAP_PROP_FPS)
while(cap.isOpened()):
    sum=0
    for i in range( 0,100 ):
        ret, frame = cap.read()
        start=time.time()
        hsv = cv2.cvtColor( frame , cv2.COLOR_BGR2HSV )

        h = frame[:,:,0]
        s = frame[:,:,1]
        v = frame[:,:,2]

        mask_g = np.zeros(h.shape, dtype=np.uint8)
        mask_g[ (h>20) & (h <100) & (s>200) & (s < 255) & (v>50) & (v<150 ) ] = 255
        end = time.time()
        sum+= (end- start)
        cv2.imshow('ReadM',mask_g )
        cv2.waitKey(1)
        if i % 10 == 0 :
            print( i )
    print( sum/100)
cap.release()

3, how to detect under conditions that can not be done with inRange with numpy

numpy2.py

import cv2
import numpy as np
import time

src = 'v4l2src device= /dev/video0 ! image/jpeg,width=1920,height=1080 !jpegdec !videoconvert ! appsink'

cap=cv2.VideoCapture(src)


W = cap.get(cv2.CAP_PROP_FRAME_WIDTH)
H = cap.get(cv2.CAP_PROP_FRAME_HEIGHT)
fps = cap.get(cv2.CAP_PROP_FPS)
while(cap.isOpened()):
    sum=0
    for i in range( 0,100 ):
        ret, frame = cap.read()
        start=time.time()
        r = frame[:,:,0]
        g = frame[:,:,1]
        b = frame[:,:,2]

        mask_g = np.zeros(r.shape, dtype=np.uint8)
        mask_g[ ( g/r> 2.0) & (g/b>2.0) ] = 255
        end = time.time()
        sum+= (end- start)
        cv2.imshow('ReadM',mask_g )
        cv2.waitKey(1)
        if i % 10 == 0 :
            print( i )
    print( sum/100)
cap.release()

When speeding up 4 and 3 with CUPY

cupy.py

import cv2
import cupy as cp
import time

src = 'v4l2src device= /dev/video0 ! image/jpeg,width=1920,height=1080 !jpegdec !videoconvert ! appsink'

cap=cv2.VideoCapture(src)


W = cap.get(cv2.CAP_PROP_FRAME_WIDTH)
H = cap.get(cv2.CAP_PROP_FRAME_HEIGHT)
fps = cap.get(cv2.CAP_PROP_FPS)
while(cap.isOpened()):
    sum=0
    for i in range( 0,100 ):
        ret, frame = cap.read()
        start=time.time()
        frame_cupy = cp.asarray( frame )
        r = frame_cupy[:,:,0]
        g = frame_cupy[:,:,1]
        b = frame_cupy[:,:,2]

        mask_g = cp.zeros(r.shape, dtype=cp.uint8)
        mask_g[ ( g/r> 2.0) & (g/b>2.0) ] = 255
        mask_gn = cp.asnumpy( mask_g )
        end = time.time()
        sum+= (end- start)
        cv2.imshow('ReadM',mask_gn )
        cv2.waitKey(1)
        if i % 10 == 0 :
            print( i )
    print( sum/100)
cap.release()