computeFeatures.py

"""
computeFeatures.py

YOUR WORKING FUNCTION for computing features

"""
import numpy as np
import cv2
import matplotlib.pyplot as plt


from squeezenet import get_squeezenet

width = 227
height = 227
channel = 3

# from cyvlfeat import sift as cysift

# you are allowed to import other Python packages above
##########################

def preprocessImage(image):
    im = cv2.resize(image, (width, height)).astype(np.float32)
    im[:,:,0] -= 103.939
    im[:,:,1] -= 116.779
    im[:,:,2] -= 123.68
    im = im.transpose((2,0,1))
    im = np.expand_dims(im, axis=0)
    return im

def computeFeatures(img, deep_features=True):
    # Inputs
    # img: 3-D numpy array of an RGB color image
    #
    # Output
    # featvect: A D-dimensional vector of the input image 'img'
    #
    #########################################################################
    # ADD YOUR CODE BELOW THIS LINE
    
    # This is the baseline method: 192-D RGB colour feature histogram
    if deep_features:
        processed_img = preprocessImage(img)
        squeezenet_model = get_squeezenet(nb_classes=10,
            path_to_weights='model/squeezenet_food.h5',
            dim_ordering='th')
        featvect = squeezenet_model.predict(processed_img)
    else:
        rhist, rbins = np.histogram(img[:,:,0], 64, normed=True)
        ghist, gbins = np.histogram(img[:,:,1], 64, normed=True)
        bhist, bbins = np.histogram(img[:,:,2], 64, normed=True)
        featvect = np.concatenate((rhist, ghist, bhist))

    # img_gray = cv2.cvtColor(img, cv2.COLOR_RGB2GRAY)

    # f2, des2 = cysift.sift(img_gray, peak_thresh=8, edge_thresh=5, compute_descriptor=True)

    # featvect = des2

    # print(des2.shape)
    
    
    # This creates a 300-D vector of random values as features!     
    #featvect = np.random.rand(300, 1)
    
    # END OF YOUR CODE
    #########################################################################
    return featvect