如何训练python函数返回所需的结果？

Question

The problem I'm looking at is that I want to detect with a fairly reasonable level of certainty whether an image is black or mostly black. I already have the code written to get the color histogram, and the next step is to write a function that will take the (r,g,b) tuple and give me a bool indicating whether it's black or close to it. It's OK for this to not be 100% accurate, but it would be better to err toward false positives.

def average_image_color(i):
    h = i.histogram()

    # split into red, green, blue
    r = h[0:256]
    g = h[256:256*2]
    b = h[256*2: 256*3]

    # perform the weighted average of each channel:
    # the *index* is the channel value, and the *value* is its weight
    return (
        sum( i*w for i, w in enumerate(r) ) / sum(r),
        sum( i*w for i, w in enumerate(g) ) / sum(g),
        sum( i*w for i, w in enumerate(b) ) / sum(b))

I have a set of test images that I can use as a corpus. What's the best library/approach to this?

The function I'd be hoping to train would be something like

def is_black(r, g, b):
    if magic_says_black():
        return True
    return False

Answer 1

Since you are only concerned about brightness, it would be easier if you converted the image to grayscale so you only have to work with one channel instead of three.

Then you have a number of options:

• image is mostly black if average pixel intensity is above an empirically determined threshold;
• count the number of pixels that are more than some threshold
• if you have plenty of example images, train a classifier such as an SVM using a grayscale histogram (this does seem like using a sledge hammer to crack a walnut). You will find plenty of classifiers in the scikit-learn package .