detect object and determine horizontal angle

asked 2018-09-18 04:26:39 -0500

I am trying to detect an object and put a red line on it, depending on its horizontal angle. Problem is that the lighting of the images changes drastically. Some of them are of good quality and some are not.

Example of good quality Image: image description

And that is how I want it to look after my program ran: image description

This is my code:

import cv2
import numpy as np

img = cv2.imread('img_in.png')
img_gs = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
img_gs = cv2.GaussianBlur(img_gs,(5,5),0)

_, thresh = cv2.threshold(img_gs,
mat = np.argwhere(thresh == 255)
mat[:, [0, 1]] = mat[:, [1, 0]]
mat = np.array(mat).astype(np.float32)
m, e = cv2.PCACompute(mat, mean = np.array([]))
center = tuple(m[0])
endpoint1 = tuple(m[0] + e[0]*100)
endpoint3 = tuple(m[0] - e[0]*100)
red_color = (0, 0, 255), center, 3, red_color)
cv2.line(img, center, endpoint1, red_color)
cv2.line(img, center, endpoint3, red_color)
cv2.imwrite('img_out.png', img)

For an image of this quality, it works. The Problem now is that I have several thousand images that differ in lighting. Two examples:

Mediocre Lighting: image description

Bad Lighting: image description

Also, there are images on which the object is brighter than the background (like the good quality example) and images on which the object is darker than the background (Bad Lighting Image). I know that I can change the threshold of every image to change the detection and that is basically what I'm doing since several hours. But the result is not very pleasing. Also I am trying to differentiate between images with an if-loop. Problem is that the only parameters I have are Max Brightness of Image Pixel, Min Brightness of Image Pixel, Mean of Image Pixels. With that I can differentiate between Images to a certain degree but there are images that are very different despite sharing these parameters. I'd be a happy man if there is a better way to do this.

edit retag flag offensive close merge delete


Did you try Ostu's method in cv2.threshold. It probably will get you better binarization result. After that you need to determine whether your object is foreground or background. I guess you could try using area if your object is relatively small.

yancey gravatar imageyancey ( 2018-09-18 14:49:42 -0500 )edit

I tried otsu's method but it didn't yield a better result. I will look into area

Artur_exe gravatar imageArtur_exe ( 2018-09-23 12:00:05 -0500 )edit