wld-feature-extractor

wld-feature-extractor.py

import cv2
import sys
import numpy as np
from scipy import signal
import matplotlib.pyplot as plt

def get_wld_feats(gray_img):
    f00 =np.array([[1, 1, 1], [1, -8, 1], [1, 1, 1]])
    f10 = np.array([[0, -1, 0], [0, 0, 0], [0, 1, 0]])
    f11 = np.array([[0, 0, 0], [1, 0, -1], [0, 0, 0]])
    f00_img = signal.convolve2d(gray_img, f00, mode='same')
    f11_img = signal.convolve2d(gray_img, f11, mode='same')
    f10_img = signal.convolve2d(gray_img, f10, mode='same')
    chi = np.arctan(f00_img/(gray_img+1e-10))
    chi_j = chi.flatten()
    theta = np.pi + np.arctan2(f11_img, (f10_img+1e-10))
    M = 6
    T = 8
    S = 4
    WLD = np.zeros((chi_j.shape[0], T))
    T_intervals = np.linspace(0, 2*np.pi, T+1)
    M_intervals = np.linspace(-np.pi/2.0, np.pi/2.0, M+1)
    h_tms = np.zeros((T, M, S))
    for i in range(T):
        t_lo = T_intervals[i]
        t_hi = T_intervals[i+1]
        for j in range(M):
            m_lo = M_intervals[j]
            m_hi = M_intervals[j+1]
            S_intervals = np.linspace(m_lo, m_hi, S + 1)
            for k in range(S):
                s_lo = S_intervals[k]
                s_hi = S_intervals[k+1]
                h_tms[i, j, k] = np.sum((chi >= s_lo) & (chi <= s_hi) & (theta >= t_lo) & (theta <= t_hi))        
    h = []
    h_img = []
    for i in range(M):
        hm = []
        for j in range(T):
            l = list(h_tms[j, i, :])
            hm += l
        h_img.append(hm)
        h += hm
    return np.array(h)
    

imgpath = sys.argv[1]
img = cv2.imread(imgpath)
gray_img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
feats = get_wld_feats(gray_img)
print feats.shape