python类imread()的实例源码

def test_image(addr):
    target = ['angry','disgust','fear','happy','sad','surprise','neutral']
    font = cv2.FONT_HERSHEY_SIMPLEX

    im = cv2.imread(addr)
    gray = cv2.cvtColor(im, cv2.COLOR_BGR2GRAY)
    faces = faceCascade.detectMultiScale(gray,scaleFactor=1.1)

    for (x, y, w, h) in faces:
            cv2.rectangle(im, (x, y), (x+w, y+h), (0, 255, 0), 2,5)
            face_crop = im[y:y+h,x:x+w]
            face_crop = cv2.resize(face_crop,(48,48))
            face_crop = cv2.cvtColor(face_crop, cv2.COLOR_BGR2GRAY)
            face_crop = face_crop.astype('float32')/255
            face_crop = np.asarray(face_crop)
            face_crop = face_crop.reshape(1, 1,face_crop.shape[0],face_crop.shape[1])
            result = target[np.argmax(model.predict(face_crop))]
            cv2.putText(im,result,(x,y), font, 1, (200,0,0), 3, cv2.LINE_AA)

    cv2.imshow('result', im)
    cv2.imwrite('result.jpg',im)
    cv2.waitKey(0)

def __init__(self, filename, folder=None, classifier=None):
        """
        :param filename: image with sudoku
        :param folder: folder where to save debug images
        :param classifier: digit classifier
        """
        self.filename = os.path.basename(filename)
        image = cv2.imread(filename)
        self.image = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
        self.folder = folder or FOLDER
        os.mkdir(os.path.join(self.folder, 'debug/'))
        self.classifier = classifier or DigitClassifier()
        # Default initial values
        self.perspective = False
        self.debug = True
        self.counter = 0
        self.step = -1

def encode(img_path, wm_path, res_path, alpha):
    img = cv2.imread(img_path)
    img_f = np.fft.fft2(img)
    height, width, channel = np.shape(img)
    watermark = cv2.imread(wm_path)
    wm_height, wm_width = watermark.shape[0], watermark.shape[1]
    x, y = range(height / 2), range(width)
    random.seed(height + width)
    random.shuffle(x)
    random.shuffle(y)
    tmp = np.zeros(img.shape)
    for i in range(height / 2):
        for j in range(width):
            if x[i] < wm_height and y[j] < wm_width:
                tmp[i][j] = watermark[x[i]][y[j]]
                tmp[height - 1 - i][width - 1 - j] = tmp[i][j]
    res_f = img_f + alpha * tmp
    res = np.fft.ifft2(res_f)
    res = np.real(res)
    cv2.imwrite(res_path, res, [int(cv2.IMWRITE_JPEG_QUALITY), 100])

def loadImgs(imgsfolder, rows, cols):
    myfiles = glob.glob(imgsfolder+'*.jpg', 0)
    nPics = len(myfiles)
    X = np.zeros((nPics, rows, cols), dtype = 'uint8')
    i = 0; imgNames = []
    for filepath in myfiles:
        sd = filepath.rfind('/'); ed = filepath.find('.'); filename = filepath[int(sd+1):int(ed)]
        imgNames.append(filename)  

        temp = cv2.imread(filepath, 0)
        if temp == None:
            continue
        elif temp.size < 1000:
            continue
        elif temp.shape == [rows, cols, 1]:
            X[i,:,:] = temp
        else:
            X[i,:,:] = cv2.resize(temp,(cols, rows), interpolation = cv2.INTER_CUBIC)
        i += 1
    return X, imgNames

def downscale(old_file_name):
    img = cv2.imread(os.path.join(old_file_name))

    new_file_name = (old_file_name
                     .replace('training', 'training_' + str(min_size))
                     .replace('validation', 'validation_' + str(min_size))
                     .replace('testing', 'testing_' + str(min_size))
                     )

    height, width, _ = img.shape

    if width > height:
        new_width = int(1.0 * width / height * min_size)
        new_height = min_size

    else:
        new_height = int(1.0 * height / width * min_size)
        new_width = min_size

    img_new = cv2.resize(img, (new_width, new_height), interpolation=cv2.INTER_LINEAR)
    cv2.imwrite(new_file_name, img_new)

def pull_item(self, index):
        img_id = self.ids[index]

        target = ET.parse(self._annopath % img_id).getroot()
        img = cv2.imread(self._imgpath % img_id)
        height, width, channels = img.shape

        if self.target_transform is not None:
            target = self.target_transform(target, width, height)

        if self.transform is not None:
            target = np.array(target)
            img, boxes, labels = self.transform(img, target[:, :4], target[:, 4])
            # to rgb
            img = img[:, :, (2, 1, 0)]
            # img = img.transpose(2, 0, 1)
            target = np.hstack((boxes, np.expand_dims(labels, axis=1)))
        return torch.from_numpy(img).permute(2, 0, 1), target, height, width
        # return torch.from_numpy(img), target, height, width

def execute_Threshold(proxy,obj):

    try: img=obj.sourceObject.Proxy.img.copy()
    except: img=cv2.imread(__dir__+'/icons/freek.png')

    # img = cv2.imread('dave.jpg',0) ??
    img = cv2.medianBlur(img,5)
    img = cv2.cvtColor(img,cv2.COLOR_BGR2GRAY)


    if obj.globalThresholding:
        ret,th1 = cv2.threshold(img,obj.param1,obj.param2,cv2.THRESH_BINARY)
        obj.Proxy.img = cv2.cvtColor(th1, cv2.COLOR_GRAY2RGB)

    if obj.adaptiveMeanTresholding:
        th2 = cv2.adaptiveThreshold(img,255,cv2.ADAPTIVE_THRESH_MEAN_C,\
                cv2.THRESH_BINARY,11,2)
        obj.Proxy.img = cv2.cvtColor(th2, cv2.COLOR_GRAY2RGB)

    if obj.adaptiveGaussianThresholding:
        th3 = cv2.adaptiveThreshold(img,255,cv2.ADAPTIVE_THRESH_GAUSSIAN_C,\
            cv2.THRESH_BINARY,17,2)
        obj.Proxy.img = cv2.cvtColor(th3, cv2.COLOR_GRAY2RGB)

def loadLogoSet(path, rows,cols,test_data_rate=0.15):
    random.seed(612)
    _, imgID = readItems('data.txt')
    y, _ = modelDict(path)
    nPics =  len(y)
    faceassset = np.zeros((nPics,rows,cols), dtype = np.uint8) ### gray images
    noImg = []
    for i in range(nPics):
        temp = cv2.imread(path +'logo/'+imgID[i]+'.jpg', 0)
        if temp == None:
            noImg.append(i)
        elif temp.size < 1000:
            noImg.append(i)
        else:
            temp = cv2.resize(temp,(cols, rows), interpolation = cv2.INTER_CUBIC)
            faceassset[i,:,:] = temp
    y = np.delete(y, noImg,0); faceassset = np.delete(faceassset, noImg, 0)
    nPics = len(y)
    index = random.sample(np.arange(nPics), int(nPics*test_data_rate))
    x_test = faceassset[index,:,:]; x_train = np.delete(faceassset, index, 0)
    y_test = y[index]; y_train = np.delete(y, index, 0)
    return (x_train, y_train), (x_test, y_test)

def test_color():
    image = cv2.imread('data/Lenna.png')
    image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)

    noise = (np.random.rand(image.shape[0], image.shape[1], 3) - 0.5) * 50
    image_noise = image + noise

    radius = [1, 2, 4]
    eps = [0.005]

    combs = list(itertools.product(radius, eps))

    vis.plot_single(to_32F(image), title='origin')
    vis.plot_single(to_32F(image_noise), title='noise')

    for r, e in combs:
        GF = GuidedFilter(image, radius=r, eps=e)
        vis.plot_single(to_32F(GF.filter(image_noise)), title='r=%d, eps=%.3f' % (r, e))

def _get_image_blob(roidb, scale_inds):
  """Builds an input blob from the images in the roidb at the specified
  scales.
  """
  num_images = len(roidb)
  processed_ims = []
  im_scales = []
  for i in range(num_images):
    im = cv2.imread(roidb[i]['image'])
    if roidb[i]['flipped']:
      im = im[:, ::-1, :]
    target_size = cfg.TRAIN.SCALES[scale_inds[i]]
    im, im_scale = prep_im_for_blob(im, cfg.PIXEL_MEANS, target_size,
                    cfg.TRAIN.MAX_SIZE)
    im_scales.append(im_scale)
    processed_ims.append(im)

  # Create a blob to hold the input images
  blob = im_list_to_blob(processed_ims)

  return blob, im_scales

def get_batch():
    ran = random.randint(600, data_size)
    #print(ran)
    image = []
    label = []
    label_0 = []
    n_pic = ran
    # print(n_pic)
    for i in range(batch_size * n_steps):
        frame_0 = cv2.imread('./cropedoriginalPixel2/%d.jpg' % (n_pic+i), 0)
        frame_0 = cv2.resize(frame_0, (LONGITUDE, LONGITUDE))
        frame_0 = np.array(frame_0).reshape(-1)
        image.append(frame_0)
        #print(np.shape(image))
    for i in range(batch_size):
        frame_1 = cv2.imread('./cropedoriginalPixel2/%d.jpg' % (n_pic + batch_size * (i+1) ), 0)
        frame_1 = cv2.resize(frame_1, (LONGITUDE, LONGITUDE))
        frame_1 = np.array(frame_1).reshape(-1)
        label.append(frame_1)
    for i in range(batch_size):
        frame_2 = cv2.imread('./cropedoriginalUS2/%d.jpg' % (n_pic + batch_size * (i+1) ), 0)
        frame_2 = cv2.resize(frame_2, (LONGITUDE, LONGITUDE))
        frame_2 = np.array(frame_2).reshape(-1)
        label_0.append(frame_2)
    return image , label , label_0

def get_batch(batch_size=20,data_size=6498):
    ran = np.random.choice(data_size, batch_size,replace=False)
    image=[]
    outline=[]
    for i in range(batch_size):
        n_pic=ran[i]
        #print(n_pic)
        frame_0 = cv2.imread('./cropPicY/%d.jpg' % n_pic,0)
        frame_0 = cv2.resize(frame_0, (24, 24))
        frame_0 = np.array(frame_0).reshape(-1)
        # print('np',frame_0)
        # frame_0 = gray2binary(frame_0)
        #print (frame_0)
        frame_1 = cv2.imread('./cropPicX/%d.jpg' % n_pic, 0)
        frame_1 = cv2.resize(frame_1, (24, 24))
        frame_1 = np.array(frame_1).reshape(-1)
        frame_1 = gray2binary(frame_1)
        image.append(frame_0)
        outline.append(frame_1)
        #print(image)
    return np.array(image),np.array(outline)

def get_train_batch(noise=0):
    ran = random.randint(600, data_size)
    #print(ran)
    image = []
    label = []
    label_0 = []
    n_pic = ran
    # print(n_pic)
    for i in range(batch_size ):
        frame_0 = cv2.imread('./cropedoriginalPixel2/%d.jpg' % (n_pic+i), 0)
        frame_0 = add_noise(frame_0, n = noise)
        frame_0 = cv2.resize(frame_0, (LONGITUDE, LONGITUDE))
        frame_0 = np.array(frame_0).reshape(-1)
        image.append(frame_0)
        #print(np.shape(image))
    for i in range(batch_size):
        frame_1 = cv2.imread('./cropedoriginalPixel2/%d.jpg' % (n_pic + batch_size * (i+1) ), 0)
        frame_1 = cv2.resize(frame_1, (LONGITUDE, LONGITUDE))
        frame_1 = np.array(frame_1).reshape(-1)
        label.append(frame_1)
    return image , label

def get_batch(batch_size=20,data_size=6498):
    ran = np.random.choice(data_size, batch_size,replace=False)
    image=[]
    for i in range(batch_size):
        n_pic=ran[i]
        #print(n_pic)
        frame_0 = cv2.imread('./cropPicX/%d.jpg' % n_pic,0)
        frame_0 = cv2.resize(frame_0, (24, 24))
        frame_0 = np.array(frame_0).reshape(-1)
        image.append(frame_0)
        #print(image)
    return np.array(image)


# Visualize decoder setting
# Parameters

def get_batch(batch_size=20,data_size=6498):
    ran = np.random.choice(data_size, batch_size,replace=False)
    image=[]
    outline=[]
    for i in range(batch_size):
        n_pic=ran[i]
        #print(n_pic)
        frame_0 = cv2.imread('./easyPixelImage2/%d.jpg' % n_pic,0)
        frame_0 = cv2.resize(frame_0, (24, 24))
        frame_0 = np.array(frame_0).reshape(-1)
        # print('np',frame_0)
        # frame_0 = gray2binary(frame_0)
        #print (frame_0)
        frame_1 = cv2.imread('./easyPixelImage2/%d.jpg' % n_pic, 0)
        frame_1 = cv2.resize(frame_1, (24, 24))
        frame_1 = np.array(frame_1).reshape(-1)
        frame_1 = gray2binary(frame_1)
        image.append(frame_0)
        outline.append(frame_1)
        #print(image)
    return np.array(image),np.array(outline)

def get_train_batch(noise=500):
    ran = np.random.randint(600,5800,size=10,dtype='int')
    #print(ran)
    image = []
    label = []
    label_0 = []
    n_pic = ran
    # print(n_pic)
    for i in range(10):
        frame_0 = cv2.imread('./cropedoriginalPixel2/%d.jpg' % (n_pic[i]), 0)
        frame_0 = add_noise(frame_0, n = noise)
        frame_0 = cv2.resize(frame_0, (24, 24))
        frame_0 = np.array(frame_0).reshape(-1)
        frame_0 = frame_0 / 255.0
        image.append(frame_0)
        #print(np.shape(image))
    for i in range(10):
        frame_1 = cv2.imread('./cropedoriginalPixel2/%d.jpg' % (n_pic[i]), 0)
        frame_1 = cv2.resize(frame_1, (24, 24))
        frame_1 = np.array(frame_1).reshape(-1)
        frame_1 = gray2binary(frame_1)
        label.append(frame_1)
    return np.array(image,dtype='float') , np.array(label,dtype='float')

def get_test_batch(noise=500):
    ran = np.random.randint(5800,6000,size=10,dtype='int')
    #print(ran)
    image = []
    label = []
    label_0 = []
    n_pic = ran
    # print(n_pic)
    for i in range(10):
        frame_0 = cv2.imread('./cropedoriginalPixel2/%d.jpg' % (n_pic[i]), 0)
        frame_0 = add_noise(frame_0, n = noise)
        frame_0 = cv2.resize(frame_0, (24, 24))
        frame_0 = np.array(frame_0).reshape(-1)
        frame_0 = frame_0 / 255.0
        image.append(frame_0)
        #print(np.shape(image))
    for i in range(10):
        frame_1 = cv2.imread('./cropedoriginalPixel2/%d.jpg' % (n_pic[i]), 0)
        frame_1 = cv2.resize(frame_1, (24, 24))
        frame_1 = np.array(frame_1).reshape(-1)
        frame_1 = gray2binary(frame_1)
        label.append(frame_1)
    return np.array(image,dtype='float') , np.array(label,dtype='float')

def get_data(datadir):
    #datadir = args.data
    # assume each image is 512x256 split to left and right
    imgs = glob.glob(os.path.join(datadir, '*.jpg'))
    data_X = np.zeros((len(imgs),3,img_cols,img_rows))
    data_Y = np.zeros((len(imgs),3,img_cols,img_rows))  
    i = 0
    for file in imgs:
        img = cv2.imread(file,cv2.IMREAD_COLOR)
        img = cv2.resize(img, (img_cols*2, img_rows)) 
        #print('{} {},{}'.format(i,np.shape(img)[0],np.shape(img)[1]))
        img = np.swapaxes(img,0,2)

        X, Y = split_input(img)

        data_X[i,:,:,:] = X
        data_Y[i,:,:,:] = Y
        i = i+1
    return data_X, data_Y

def gen_image_frames(img_path, fps, duration):
    """Generate frames from a image so that can play like a video.

    Parameters
    ----------
    img_path : string
        absolute path to the image.
    fps : int
        frame rates per second
    duration : float
        duration of the sequence in second
    """
    num_frames = int(fps*duration)

    img = cv2.imread(img_path)
    frames = []
    for i in xrange(num_frames):
        frames.append(img)

    return frames, len(frames)

def _get_image_blob(roidb, scale_inds):
    """Builds an input blob from the images in the roidb at the specified
    scales.
    """
    num_images = len(roidb)
    processed_ims = []
    im_scales = []
    im_shapes = np.zeros((0, 2), dtype=np.float32)
    for i in xrange(num_images):
        im = cv2.imread(roidb[i]['image'])
        if roidb[i]['flipped']:
            im = im[:, ::-1, :]
        target_size = cfg.TRAIN.SCALES[scale_inds[i]]
        im, im_scale, im_shape = prep_im_for_blob(im, cfg.PIXEL_MEANS, target_size)
        im_scales.append(im_scale)
        processed_ims.append(im)
        im_shapes = np.vstack((im_shapes, im_shape))

    # Create a blob to hold the input images
    blob = im_list_to_blob(processed_ims)

    return blob, im_scales, im_shapes

def get_image_features(self, img_file, stride=5, padding=True):
        """
        Take an image file as input, and output an array of image features whose matrix size is
        based on the image size. When no padding, and the image size is smaller than the required
        feature space size (in x or y direction), the image is not checked, and this method will
        return a tuple of two empty lists; When padding is True, and the image size is more than
        4 pixels smaller than the require feature space size (in x or y direction), the image is
        not checked either. This method can be used by both the trainer and predictor.
        Args:
            img_file: The file name of the image.
            stride: Optional. The stride of the sliding.
            padding: Optional. Whether to pad the image to fit the feature space size or to
                discard the extra pixels if padding is False.
        Returns:
            coordinates: A list of coordinates, each of which contains y and x that are the top
                left corner offsets of the sliding window.
            features: A matrix (python list), in which each row contains the features of the
                sampling sliding window, while the number of rows depends on the image size of
                the input.
        """
        img = cv2.imread(img_file)
        img_arr = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)

        return self.get_image_array_features(img_arr, stride, padding)

def s1_predict(config_file, model_dir, model_file, predict_file_list, out_dir):
    """
    This function serves as a test/validation tool during the model development. It is not used as
    a final product in part of the pipeline.
    """
    with open(config_file) as config_buffer:
        config = json.loads(config_buffer.read())

    with tf.Graph().as_default() as graph:
        converted_model = ConvertedModel(config, graph, 's1_keras', model_dir, model_file)

    with tf.Session(graph=graph) as sess:
        for img_file in predict_file_list:
            image = cv2.imread(img_file)
            boxes = converted_model.predict(sess, image)
            image = draw_boxes(image, boxes)

            _, filename = os.path.split(img_file)
            cv2.imwrite(os.path.join(out_dir, filename), image)

def get_batch_idx(self, idx):
        hh = self.inp_height
        ww = self.inp_width
        x = np.zeros([len(idx), hh, ww, 3], dtype='float32')
        orig_height = []
        orig_width = []
        ids = []
        for kk, ii in enumerate(idx):
            fname = self.ids[ii]
            ids.append('{:06}'.format(ii))
            x_ = cv2.imread(fname).astype('float32') / 255
            x[kk] = cv2.resize(
                x_, (self.inp_width, self.inp_height),
                interpolation=cv2.INTER_CUBIC)
            orig_height.append(x_.shape[0])
            orig_width.append(x_.shape[1])
            pass
        return {
            'x': x,
            'orig_height': np.array(orig_height),
            'orig_width': np.array(orig_width),
            'id': ids
        }

def process_frame(frame_number, frame, keypoint_data, detector, matcher):
    log = logging.getLogger("process_frame")

    # Create a copy of source frame to draw into
    processed = frame.copy()

    gray_frame = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
    kp, des = detector.detectAndCompute(frame, None)

    # Match descriptors
    matches = matcher.match(keypoint_data.descriptors, des)

    # Sort them in order of distance
    matches = sorted(matches, key = lambda x:x.distance)

    processed = drawMatches(cv2.imread('car.png',0), keypoint_data.keypoints, gray_frame, kp, matches[:])

    return processed

# ============================================================================

def check_image(name):
    expected_data = json.loads(open('./img/' + name + '.json').read())
    if not expected_data['enabled']:
        return

    expected_targets = expected_data['targets']

    img = cv2.imread('./img/' + name + '.jpg', cv2.IMREAD_COLOR)
    hsv = cv2.cvtColor(img, cv2.COLOR_BGR2HSV)
    args = config.copy()
    args['img'] = hsv
    args['output_images'] = {}

    actual_targets = find(**args)

    # make sure same number of targets are detected
    assert len(expected_targets) == len(actual_targets)

    # targets is a list of 2-tuples with expected and actual results
    targets = zip(expected_targets, actual_targets)
    # compare all the different features of targets to make sure they match
    for pair in targets:
        expected, actual = pair
        # make sure that the targets are close to where they are supposed to be
        assert is_close(expected['pos']['x'], actual['pos']['x'], 0.02)
        assert is_close(expected['pos']['y'], actual['pos']['y'], 0.02)
        # make sure that the targets are close to the size they are supposed to be
        assert is_close(expected['size']['width'], actual['size']['width'], 0.02)
        assert is_close(expected['size']['height'], actual['size']['height'], 0.02)

def create_composite_image_coin_id(coin_id, crop_dir, data_dir):
    images = []
    images_gif = []

    for id in range(0,56):
        image_id = coin_id * 100 + id
        crop = ci.get_rotated_crop(crop_dir, image_id, 56, 0)
        images.append(crop)
        filename =  ci.get_filename_from(image_id,crop_dir)
        images_gif.append(imageio.imread(filename))

    composite_image = ci.get_composite_image(images, 8, 8)
    cv2.imwrite(data_dir + str(coin_id) + '.png', composite_image)
    imageio.mimsave(data_dir + str(coin_id) + '.gif', images_gif)

    return

def get_image_array(roidb, scales, scale_indexes, need_mean=True):
    """
    build image array from specific roidb
    :param roidb: images to be processed
    :param scales: scale list
    :param scale_indexes: indexes
    :return: array [b, c, h, w], list of scales
    """
    num_images = len(roidb)
    processed_ims = []
    im_scales = []
    for i in range(num_images):
        im = cv2.imread(roidb[i]['image'])
        if roidb[i]['flipped']:
            im = im[:, ::-1, :]
        target_size = scales[scale_indexes[i]]
        im, im_scale = image_processing.resize(im, target_size, config.MAX_SIZE)
        im_tensor = image_processing.transform(im, config.PIXEL_MEANS, need_mean=need_mean)
        processed_ims.append(im_tensor)
        im_scales.append(im_scale)
    array = image_processing.tensor_vstack(processed_ims)
    return array, im_scales

def _get_image_blob(roidb, scale_inds):
    """Builds an input blob from the images in the roidb at the specified
    scales.
    """
    num_images = len(roidb)
    processed_ims = []
    im_scales = []
    for i in xrange(num_images):
        im = cv2.imread(roidb[i]['image'])
        if roidb[i]['flipped']:
            im = im[:, ::-1, :]
        target_size = cfg.TRAIN.SCALES[scale_inds[i]]
        im, im_scale = prep_im_for_blob(im, cfg.PIXEL_MEANS, target_size,
                                        cfg.TRAIN.MAX_SIZE)
        im_scales.append(im_scale)
        processed_ims.append(im)

    # Create a blob to hold the input images
    blob = im_list_to_blob(processed_ims)

    return blob, im_scales

def imdb_proposals(net, imdb):
    """Generate RPN proposals on all images in an imdb."""

    _t = Timer()
    imdb_boxes = [[] for _ in xrange(imdb.num_images)]
    for i in xrange(imdb.num_images):
        im = cv2.imread(imdb.image_path_at(i))
        _t.tic()
        imdb_boxes[i], scores = im_proposals(net, im)
        _t.toc()
        print 'im_proposals: {:d}/{:d} {:.3f}s' \
              .format(i + 1, imdb.num_images, _t.average_time)
        if 0:
            dets = np.hstack((imdb_boxes[i], scores))
            # from IPython import embed; embed()
            _vis_proposals(im, dets[:3, :], thresh=0.9)
            plt.show()

    return imdb_boxes

def _get_feature_scale(self, num_images=100):
        TARGET_NORM = 20.0 # Magic value from traditional R-CNN
        _t = Timer()
        roidb = self.imdb.roidb
        total_norm = 0.0
        count = 0.0
        inds = npr.choice(xrange(self.imdb.num_images), size=num_images,
                          replace=False)
        for i_, i in enumerate(inds):
            im = cv2.imread(self.imdb.image_path_at(i))
            if roidb[i]['flipped']:
                im = im[:, ::-1, :]
            _t.tic()
            scores, boxes = im_detect(self.net, im, roidb[i]['boxes'])
            _t.toc()
            feat = self.net.blobs[self.layer].data
            total_norm += np.sqrt((feat ** 2).sum(axis=1)).sum()
            count += feat.shape[0]
            print('{}/{}: avg feature norm: {:.3f}'.format(i_ + 1, num_images,
                                                           total_norm / count))

        return TARGET_NORM * 1.0 / (total_norm / count)