Identify object on a conveyor belt

edit

Binarize the image with color or greyscale thresholding and then do a cv::findContours(). Loop through contours filtering by contour size. For those contours within the accepted size bounds, compare moments, either Hu or Flusser, to the canonical item, maybe using Mahalnobis distance for a conscious false negative rate.

Thank you! I will start with your advice! Will I need to put a glare filter on the camera lens? Is raspberry pi camera module a good choice?

Try to make your lighting more diffuse, there will be much less glare if there is ambient light and not a point light source. A single color highly saturated background would likely help as well, in which case you could try HSV thresholding.

I'm getting really close now!

I've changed the background to a bright orange.

My code is as follows:

-Take an image -Convert image from BGR to HSV -Threshold the image to filter out the orange -FindContours -Filter out the contours that doesnt match my wanted area. -Process the Moments and HuMoments of the contours that are left -Calculate the centroid from the Moments -Draw the contours and centroid in the original image

In the image I have 2 objects that are with their faces down and 2 object are face up. I only want the program to recognize the 2 object faced upwards. Im trying to print the Moments and HuMoments Array but dont know how to filter out values so that the face-down ones will be filtered out?

area_min=3000 area_max=4000

for i in range(len(contours)):
    cnt=contours[i]
    area=cv2.contourArea(cnt)
    centres=[]
    if (area > area_min) and (area < area_max):
            moments=cv2.moments(contours[i])
            hu=cv2.HuMoments(moments)
            centres.append((int(moments['m10']/moments['m00']),int(moments['m01']/moments['m00'])))
            cv2.circle(image, centres[-1],3,(0,0,255),-1)
            cv2.drawContours(image, contours, i, (0,255,0), 2)

            #print(centres)
            print(moments)

Resulting image: link text

The returned moments values are too long to attach. And also the HuMoments values.

The sign of the final Hu moment can discriminate mirror images of contours (face-down vs. face-up in your case). Also, have a look at cv::Mahalanobis()for obtaining a probability (0-1) that a contour belongs to canonical class.