"Not Enough Image Data" Error when Converting CV2 Numpy Array into RGB Image

Question

In the code below, I am attempting to output a single face (cropped from a larger image) with CV2:

def machine_pst():
    mlimg = request.files.get("mlimg")
    fname = mlimg.filename
    filepath = "/home/assets/faces/"
    mlimg.save(filepath + fname, overwrite = True)
    full_path = filepath + fname
    cascPath = "/home/assets/haarcascade_frontalface_default.xml"
    detector = cv2.CascadeClassifier(cascPath)
    faceSamples=[]
    pilImage=Image.open(full_path).convert('L')
    imageNp=np.array(pilImage,'uint8')
    faces=detector.detectMultiScale(imageNp)
    for (x,y,w,h) in faces:
        faceSamples.append(imageNp[y:y+h,x:x+w])
    img = Image.fromarray(faceSamples[0], 'RGB')

   cv2.imwrite("/home/assets/faces/read.png", img)
   source = "/static/faces/read.png"
   return template("home/machineout", source = source)

With source being passed as a parameter into img src="{{source}}

If I return the length of faces in an image with 3 faces, I get "3", so that seems to work nicely and if I return any index of faceSamples (e.g. faceSamples[0]), I get data returned as well, but when I try to turn that face sample into an image using ...

img = Image.fromarray(faceSamples[0], 'RGB')

I get a ValueError that there is "not enough image data"

I understand (from a previous answer) that detectMultiScale returns rectangles, not images, but with my additional Numpy code, is that still the case? Am I still not fully understanding what the faceSamples array is returning? Can this not be directly turned back into an RGB image with the last snippet of code?

Answer 1

Your problem is here:

pilImage=Image.open(full_path).convert('L')
imageNp=np.array(pilImage,'uint8')

That is, you converted imageNp into a single channel, gray image. Then it makes little sense to do

img = Image.fromarray(faceSamples[0], 'RGB')

as faceSamples[0] is also a gray image.

Also, like @MarkSetchell's comment, you can use cv2.imread and other functions instead of PIL. They are more compatible with other openCV functions.