feature_extractor.py

import cv2
# import fhog_feature
# import cn_feature
import numpy as np
import conv_reg_config


class FeatureExtractor(object):

    def __init__(self):
        self._resolution = 1.0
        self._channel_num = 1

    def _extract_feature(self, image):
        pass

    def extract_multiple_features(self, input_images):
        features = []
        for image in input_images:
            feature = self._extract_feature(image)
            features.append(feature[np.newaxis, :, :, :])
        assert len(features) > 0

        conca_features = np.concatenate(features, axis=0)
        return conca_features

    def get_resolution(self):
        return self._resolution

    def get_channel_num(self):
        return self._channel_num


class GrayExtractor(FeatureExtractor):

    def _extract_feature(self, input_image):
        gray = cv2.cvtColor(input_image, cv2.COLOR_RGB2GRAY)
        return gray


class RgbExtractor(FeatureExtractor):

    def __init__(self):
        super(RgbExtractor, self).__init__()
        self._channel_num = 3

    def _extract_feature(self, input_image):
        return input_image


# class FhogExtractor(FeatureExtractor):
#
#     def __init__(self):
#         super(FhogExtractor, self).__init__()
#         self.cell_size = conv_reg_config.FhogCfg.CELL_SIZE
#         self.bin_num = conv_reg_config.FhogCfg.BIN_NUM
#         self._resolution = self.cell_size
#         self._channel_num = 3*self.bin_num + 4
#
#     def _extract_feature(self, input_image):
#         assert input_image.shape[2] == 3
#         im = cv2.cvtColor(input_image, cv2.COLOR_BGR2GRAY)
#         im = np.asarray(im, dtype=np.float32)
#         return fhog_feature.extract(im, bin_size=self.cell_size, n_orients=self.bin_num)

#
# class FhogCnExtractor(FeatureExtractor):
#
#     def __init__(self):
#         super(FhogCnExtractor, self).__init__()
#         self.cell_size = conv_reg_config.FhogCnCfg.CELL_SIZE
#         self.bin_num = conv_reg_config.FhogCnCfg.BIN_NUM
#         self._resolution = self.cell_size
#         self._channel_num = 3*self.bin_num + 4 + 10
#
#     def _extract_feature(self, input_image):
#         assert input_image.shape[2] == 3
#         im = cv2.cvtColor(input_image, cv2.COLOR_BGR2GRAY)
#         im = np.asarray(im, dtype=np.float32)
#         fhog = fhog_feature.extract(im, bin_size=self.cell_size, n_orients=self.bin_num)
#
#         rescaled_im = cv2.resize(input_image, (fhog.shape[1], fhog.shape[0]))
#         cn = cn_feature.extract(rescaled_im)
#
#         res = np.concatenate((fhog, cn), axis=2)
#         return res