Tensorflow实现二次元图像的超分辨率

import tensorflow as tf from configs import * from os.path import join def batch_queue_for_training(data_path): filename_queue = tf.train.string_input_producer(tf.train.match_filenames_once(join(data_path, '*.png'))) file_reader = tf.WholeFileReader() _, image_file = file_reader.read(filename_queue) patch = tf.image.decode_png(image_file, NUM_CHENNELS) # we must set the shape of the image before making batches patch.set_shape([PATCH_SIZE, PATCH_SIZE, NUM_CHENNELS]) patch = tf.image.convert_image_dtype(patch, dtype=tf.float32) if MAX_RANDOM_BRIGHTNESS > 0: patch = tf.image.random_brightness(patch, MAX_RANDOM_BRIGHTNESS) if len(RANDOM_CONTRAST_RANGE) == 2: patch = tf.image.random_contrast(patch, *RANDOM_CONTRAST_RANGE) patch = tf.image.random_flip_left_right(patch) high_res_patch = tf.image.random_flip_up_down(patch) crop_margin = PATCH_SIZE - LABEL_SIZE assert crop_margin >= 0 if crop_margin > 1: high_res_patch = tf.random_crop(patch, [LABEL_SIZE, LABEL_SIZE, NUM_CHENNELS]) # crop_pos = tf.random_uniform([2], 0, crop_margin, dtype=tf.int32) # offset = tf.convert_to_tensor([crop_pos[0], crop_pos[1], 0]) # size = tf.convert_to_tensor([CROP_SIZE, CROP_SIZE, NUM_CHENNELS]) # high_res_patch = tf.slice(patch, offset, size) # additional 1px shifting to low_res_patch, reducing the even/odd issue in nearest neighbor scaler. # shift1px = tf.random_uniform([2], -1, 2, dtype=tf.int32) # offset += tf.convert_to_tensor([shift1px[0], shift1px[1], 0]) # offset = tf.clip_by_value(offset, 0, crop_margin-1) # low_res_patch = tf.slice(patch, offset, size) downscale_size = [INPUT_SIZE, INPUT_SIZE] resize_nn = lambda: tf.image.resize_nearest_neighbor([high_res_patch], downscale_size, True) resize_area = lambda: tf.image.resize_area([high_res_patch], downscale_size, True) resize_cubic = lambda: tf.image.resize_bicubic([high_res_patch], downscale_size, True) r = tf.random_uniform([], 0, 3, dtype=tf.int32) low_res_patch = tf.case({tf.equal(r, 0): resize_nn, tf.equal(r, 1): resize_area}, default=resize_cubic)[0] # add jpeg noise to low_res_patch if JPEG_NOISE_LEVEL > 0: low_res_patch = tf.image.convert_image_dtype(low_res_patch, dtype=tf.uint8, saturate=True) jpeg_quality = 100 - 5 * JPEG_NOISE_LEVEL jpeg_code = tf.image.encode_jpeg(low_res_patch, quality=jpeg_quality) low_res_patch = tf.image.decode_jpeg(jpeg_code) low_res_patch = tf.image.convert_image_dtype(low_res_patch, dtype=tf.float32) # we must set tensor's shape before doing following processes low_res_patch.set_shape([INPUT_SIZE, INPUT_SIZE, NUM_CHENNELS]) # add noise to low_res_patch if GAUSSIAN_NOISE_STD > 0: low_res_patch += tf.random_normal(low_res_patch.get_shape(), stddev=GAUSSIAN_NOISE_STD) low_res_patch = tf.clip_by_value(low_res_patch, 0, 1.0) high_res_patch = tf.clip_by_value(high_res_patch, 0, 1.0) # low_res_patch -= 0.5 # approximate mean-zero data # high_res_patch -= 0.5 # low_res_patch = tf.clip_by_value(low_res_patch, -0.5, 0.5) # high_res_patch = tf.clip_by_value(high_res_patch, -0.5, 0.5) # Generate batch low_res_batch, high_res_batch = tf.train.shuffle_batch( [low_res_patch, high_res_patch], batch_size=BATCH_SIZE, num_threads=NUM_PROCESS_THREADS, capacity=MIN_QUEUE_EXAMPLES + 3 * BATCH_SIZE, min_after_dequeue=MIN_QUEUE_EXAMPLES) return low_res_batch, high_res_batch def batch_queue_for_testing(data_path): filename_queue = tf.train.string_input_producer(tf.train.match_filenames_once(join(data_path, '*.png'))) file_reader = tf.WholeFileReader() _, image_file = file_reader.read(filename_queue) patch = tf.image.decode_png(image_file, NUM_CHENNELS) # we must set the shape of the image before making batches patch.set_shape([PATCH_SIZE, PATCH_SIZE, NUM_CHENNELS]) patch = tf.image.convert_image_dtype(patch, dtype=tf.float32) crop_margin = PATCH_SIZE - LABEL_SIZE offset = tf.convert_to_tensor([crop_margin // 2, crop_margin // 2, 0]) size = tf.convert_to_tensor([LABEL_SIZE, LABEL_SIZE, NUM_CHENNELS]) high_res_patch = tf.slice(patch, offset, size) downscale_size = [INPUT_SIZE, INPUT_SIZE] resize_nn = lambda: tf.image.resize_nearest_neighbor([high_res_patch], downscale_size, True) resize_area = lambda: tf.image.resize_area([high_res_patch], downscale_size, True) resize_cubic = lambda: tf.image.resize_bicubic([high_res_patch], downscale_size, True) r = tf.random_uniform([], 0, 3, dtype=tf.int32) low_res_patch = tf.case({tf.equal(r, 0): resize_nn, tf.equal(r, 1): resize_area}, default=resize_cubic)[0] # add jpeg noise to low_res_patch if JPEG_NOISE_LEVEL > 0: low_res_patch = tf.image.convert_image_dtype(low_res_patch, dtype=tf.uint8, saturate=True) jpeg_quality = 100 - 5 * JPEG_NOISE_LEVEL jpeg_code = tf.image.encode_jpeg(low_res_patch, quality=jpeg_quality) low_res_patch = tf.image.decode_jpeg(jpeg_code) low_res_patch = tf.image.convert_image_dtype(low_res_patch, dtype=tf.float32) # we must set tensor's shape before doing following processes low_res_patch.set_shape([INPUT_SIZE, INPUT_SIZE, NUM_CHENNELS]) # add noise to low_res_patch if GAUSSIAN_NOISE_STD > 0: low_res_patch += tf.random_normal(low_res_patch.get_shape(), stddev=GAUSSIAN_NOISE_STD) low_res_patch = tf.clip_by_value(low_res_patch, 0, 1.0) high_res_patch = tf.clip_by_value(high_res_patch, 0, 1.0) # low_res_patch -= 0.5 # approximate mean-zero data # high_res_patch -= 0.5 # low_res_patch = tf.clip_by_value(low_res_patch, -0.5, 0.5) # high_res_patch = tf.clip_by_value(high_res_patch, -0.5, 0.5) # Generate batch low_res_batch, high_res_batch = tf.train.batch( [low_res_patch, high_res_patch], batch_size=BATCH_SIZE, num_threads=1, capacity=MIN_QUEUE_EXAMPLES + 3 * BATCH_SIZE) return low_res_batch, high_res_batch # def batch_queue_for_pair_training(data_path): # filename_queue = tf.train.string_input_producer(tf.train.match_filenames_once(join(data_path, '*.png'))) # file_reader = tf.WholeFileReader() # _, image_file = file_reader.read(filename_queue) # patch_pair = tf.image.decode_png(image_file, NUM_CHENNELS) # # we must set the shape of the image before making batches # patch_pair.set_shape([PATCH_SIZE, PATCH_SIZE * 2, NUM_CHENNELS]) # # # random brightness and contrast synchronously # # patch_pair = tf.cast(patch_pair, tf.float32) / 127.5 - 1.0 # patch_pair = tf.image.convert_image_dtype(patch_pair, dtype=tf.float32) # argumented_image = tf.image.random_brightness(patch_pair, MAX_RANDOM_BRIGHTNESS) # argumented_image = tf.image.random_contrast(argumented_image, *RANDOM_CONTRAST_RANGE) # # # patch_pair is formated like: low_res_patch | high_res_patch # low_res_patch = argumented_image[:, :PATCH_SIZE, ...] # high_res_patch = argumented_image[:, PATCH_SIZE:, ...] # # # add jpeg noise to low_res_patch # if JPEG_NOISE_LEVEL > 0: # low_res_patch = tf.image.convert_image_dtype(low_res_patch, dtype=tf.uint8, saturate=True) # jpeg_quality = 100 - 5 * JPEG_NOISE_LEVEL # jpeg_code = tf.image.encode_jpeg(low_res_patch, quality=jpeg_quality) # low_res_patch = tf.image.decode_jpeg(jpeg_code) # low_res_patch = tf.image.convert_image_dtype(low_res_patch, dtype=tf.float32) # # crop_margin = PATCH_SIZE - INPUT_SIZE # if crop_margin > 1: # # random crop synchronously # crop_pos = tf.random_uniform([2], 0, crop_margin, dtype=tf.int32) #