>
def DCT(S, N, U):    
    Pi = math.pi    
    F2 = lambda Sl:Sl[0] * math.cos(float(U) * Pi *(Sl[1]+0.5)/float(N))
    s = sum(map(F2, zip(S,range(N))))
 
    if U == 0:
        Cu = 1.0/math.sqrt(8);
    else:
        Cu = 0.5;                   
    return Cu*s
 
def DCT2(S, N):
    # start by do DCT on row    
    res = map( lambda x: [x]*N, [0]*N )
    for u in range(N):
        for v in range(N):
            res[u][v] = DCT(S[u], N, v)
 
    #redo it again on column, we will get res2 as DCT2 matrix
    res2 = map( lambda x: [x]*N, [0]*N)
    for u in range(N):
        for v in range(N):
            r = getcol(res, N, u)
            res2[v][u] = DCT(r, N, v)
 
    return res2[:]
</blockcode>

from numpy import *
from numpy.random import *
import math
import time
 
def getcol(S,col):
	return list(array(S)[:,col])
 
def DCT(S, N, U):    
	Pi = math.pi    
	F2 = lambda Sl:Sl[0] * math.cos(float(U) * Pi *(Sl[1]+0.5)/float(N))
	s = sum(map(F2, zip(S,range(N))))
 
	if U == 0:
		Cu = 1.0/math.sqrt(8);
	else:
		Cu = 0.5;                   
	return Cu*s
 
def DCT2(S, N):
	# start by do DCT on row    
	res = map( lambda x: [x]*N, [0]*N )
	for u in range(N):
		for v in range(N):
			res[u][v] = DCT(S[u], N, v)
 
	#redo it again on column, we will get res2 as DCT2 matrix
	res2 = map( lambda x: [x]*N, [0]*N)
	for u in range(N):
		for v in range(N):
			r = getcol(res, u)
			res2[v][u] = DCT(r, N, v)
 
	return res2[:]
 
s = randint(1, 256, (40,40))
begin = time.time()
DCT2(s,40)
print time.time()-begin

from numpy import *
from numpy.random import *
import time
 
def dct(s):
	row = s.shape[0]
	col = s.shape[1]
	res = zeros_like(s)
 
	cu_row = zeros_like(s)+0.5
	cu_row[:,0] = 1.0/8**0.5
	cu_col = zeros_like(s)+0.5
	cu_col[0,:] = 1.0/8**0.5
	cu_mat = cu_row*cu_col
 
	row_range = arange(row)
	col_range = arange(col)
 
	s = matrix(s)
	for i in row_range :
		for j in col_range :
			cos_row = matrix(cos(pi/row*(row_range+0.5)*i))
			cos_col = matrix(cos(pi/col*(col_range+0.5)*j)).T
			res[i][j] = cos_row * (s * cos_col)
 
	res = cu_mat*res
	return res