[PYTHON] Image processing 100 knocks Q9, Q10 (filter) speedup

Recently I started knocking 100 images. https://qiita.com/yoyoyo_/items/2ef53f47f87dcf5d1e14

Recently, I was a little interested in image processing and wanted to use numpy well, so I am very grateful for this collection of problems, but I am dissatisfied with the answer examples of Q.9. Gaussian filter and Q.10. Median filter. There was (I don't want to use the for statement in python!), So I wrote a process that uses numpy as much as possible and compared the speed.

The code I wrote this time:

def gaussian_filter(_img, filter_matrix):
    # assume that fY and fX is odd number
    (fY, fX) = filter_matrix.shape
    mY = (fY - 1) // 2
    mX = (fX - 1) // 2

    (Y, X, C) = _img.shape
    img = np.zeros((Y + mY*2, X + mX*2, C))
    img[mY:Y+mY, mX:X+mX, :] = _img

    out_img = np.zeros_like(_img, np.float32)
    for dy in range (fY):
        for dx in range(fX):
            out_img += filter_matrix[dy][dx] * img[dy:Y+dy, dx:X+dx, :]
    return out_img.astype(np.uint8)

By turning y, x of filter_matrix with for instead of y, x of the image, the addition in the image can use for of numpy. Speed comparison We called gaussian_filter 100 times and compared how many seconds it took. Since the time command is used, the time for one image load is included, but I don't think it will affect many people.

Method	time
This code	0.72
Model answer	50.60

Isn't it overwhelming? Our army is ~ ~ ~ Q10. Regarding the median filter, I think that it can be speeded up by creating an image (?) Of H * W * (filter_size) and then doing np.median (axis = 2) on that image. .. ~~ It's done.

def median_filter(_img, filter_size):
    # assume that filter_size is odd
    half_size = (filter_size - 1) // 2
    (Y, X, C) = _img.shape
    img = np.zeros((Y + half_size*2, X + half_size*2, C))
    img[half_size:Y+half_size, half_size:X+half_size, :] = _img

    out_img = np.zeros_like(_img)
    for c in range(C):
        big_ch_img = np.zeros((Y, X, filter_size**2))
        for fy in range(filter_size):
            for fx in range(filter_size):
                ch = fy * filter_size + fx
                big_ch_img[:,:,ch] = img[fy:fy+Y, fx:fx+X,c]
        out_img[:,:,c] = np.median(big_ch_img, axis=2)
    return out_img

The measurement time was 1.71 seconds vs 302.98 seconds.

By the way, I haven't come up with a way to avoid using the for statement for Pooling, so please let me know if anyone knows.