saliency.py 文件源码

def features(image, channel, levels=9, start_size=(1983, 1088), ):
    """
        Extracts features by down-scaling the image levels times,
        transforms the image by applying the function channel to 
        each scaled version and computing the difference between 
        the scaled, transformed versions.
            image :         the image
            channel :       a function which transforms the image into
                            another image of the same size
            levels :        number of scaling levels
            start_size :    tuple.  The size of the biggest image in 
                            the scaling pyramid.  The image is first
                            scaled to that size and then scaled by half
                            levels times.  Therefore, both entries in
                            start_size must be divisible by 2^levels.
    """
    image = channel(image)
    if image.shape != start_size:
        image = cv2.resize(image, dsize=start_size)

    scales = [image]
    for l in xrange(levels - 1):
        logger.debug("scaling at level %d", l)
        scales.append(cv2.pyrDown(scales[-1]))

    features = []
    for i in xrange(1, levels - 5):
        big = scales[i]
        for j in (3,4):
            logger.debug("computing features for levels %d and %d", i, i + j)
            small = scales[i + j]
            srcsize = small.shape[1],small.shape[0]
            dstsize = big.shape[1],big.shape[0]
            logger.debug("Shape source: %s, Shape target :%s", srcsize, dstsize)
            scaled = cv2.resize(src=small, dsize=dstsize)
            features.append(((i+1,j+1),cv2.absdiff(big, scaled)))

    return features