Numpythonic way for applying convolution kernel

Question

I would like to know if there is another way of applying the kernel using some non-standard numpy function.

Complexity is not going to be reduced since the operation involves applying to an image with size N x M a kernel of size K x K. So this is O(K² x N x M). Despite this maybe some time reduction by a factor could be achieved eliminating for loops.

    def apply_gaussian_kernel(img, kernel_size, max_percentage = 0.2):
        """Application of the gaussian kernel to an RGB image
            max_percentage : percentage of the maximum value of any pixel on any channel
                to define standard deviation
        """
        
        stdev = np.max(img)*max_percentage
        shape = img.shape
        N,M,RGB = shape
        
        kernel = gaussian_kernel(kernel_size)
        padding = kernel.shape[0]//2  #assuming odd sized squared kernel
        img = apply_padding(img, padding)
        
        output = np.zeros(shape=(N,M,RGB))
        
        for n in range(N):
            for m in range(M):
                for rgb in range(RGB):
                    nk = n+padding
                    mk = m+padding
                    result = np.sum(kernel * img[nk-padding:nk+padding+1,mk-padding:mk+padding+1,rgb])
                    output[n][m][rgb] = result
                                
        return output