def from_gdal_dataset(cls, dataset, band):
"""
Creates a new 1-band pyvips.Image from `dataset` at `band`
dataset: GDAL Dataset
band: Number of the band, starting from 1
"""
with LibVips.disable_warnings():
filename = dataset.GetFileList()[0]
image1 = Image.new_from_file(filename)
# Extract the band
image2 = image1.extract_band((band - 1), n=1)
# Cast to the right datatype, if necessary
datatype = dataset.GetRasterBand(band).NumPyDataType
if VImageAdapter(image2).NumPyType() == datatype:
return image2
types = dict((v, k) for k, v in cls.NUMPY_TYPES.items())
image3 = Image.new_from_memory(image2.write_to_memory(),
width=image2.width,
height=image2.height,
bands=1,
format=types[datatype])
return image3
def from_gdal_dataset(cls, dataset, band):
"""
Creates a new 1-band pyvips.Image from `dataset` at `band`
dataset: GDAL Dataset
band: Number of the band, starting from 1
"""
with LibVips.disable_warnings():
filename = dataset.GetFileList()[0]
image1 = Image.new_from_file(filename)
# Extract the band
image2 = image1.extract_band((band - 1), n=1)
# Cast to the right datatype, if necessary
datatype = dataset.GetRasterBand(band).NumPyDataType
if VImageAdapter(image2).NumPyType() == datatype:
return image2
types = dict((v, k) for k, v in cls.NUMPY_TYPES.items())
image3 = Image.new_from_memory(image2.write_to_memory(),
width=image2.width,
height=image2.height,
bands=1, format=types[datatype])
image3._buf = image2
return image3
def numpy_to_vips(np_array: np.ndarray,
width: Optional[int] = None,
height: Optional[int] = None,
n_channels: Optional[int] = None) -> VIPSImage:
"""
Convert a Numpy array to a VIPS image.
Parameters
----------
np_array : array-like
Numpy array to convert.
If 1D, it is expected it contains flattened image data.
width : int (optional)
Width of the image, must be given if `np_array` is 1D,
otherwise inferred from shape.
height : int (optional)
Height of the image, must be given if `np_array` is 1D,
otherwise inferred from shape.
n_channels : int (optional)
n_channels of the image, must be given if `np_array` is 1D,
otherwise inferred from shape.
Returns
-------
image
VIPS image representation of the array
Raises
------
ValueError
If it is impossible to convert provided array.
"""
if not np_array.flags['C_CONTIGUOUS']:
np_array = np.ascontiguousarray(np_array)
if np_array.ndim > 3:
raise NotImplementedError
elif np_array.ndim > 1:
if np_array.ndim == 2:
height_, width_ = np_array.shape
n_channels_ = 1
else:
height_, width_, n_channels_ = np_array.shape
width = width if width is not None else width_
height = height if height is not None else height_
n_channels = n_channels if n_channels is not None else n_channels_
if width * height * n_channels != np_array.size:
raise ValueError(f"Cannot convert {np_array} to VIPS image")
flat = np_array.reshape(np_array.size)
vips_format = dtype_to_vips_format[str(np_array.dtype)]
return VIPSImage.new_from_memory(flat.data, width, height, n_channels,
vips_format)
def write_buffer(self, image, resolution):
if VImageAdapter(
image).BufferSize() >= self.IMAGE_BUFFER_DISK_THRESHOLD:
logger.debug('Buffering resolution {0} to disk'.format(resolution))
vipsfile = Image.new_temp_file("%s.v")
tempfile_image = image.write(vipsfile)
return tempfile_image
logger.debug('Buffering resolution {0} to memory'.format(resolution))
return Image.new_from_memory(image.write_to_memory(), image.width,
image.height, image.bands, 'uchar')
def from_numpy_array(cls, array, width, height, bands, format):
"""
Returns a new pyvips.Image created from a NumPy `array` of pixel data.
array: The NumPy array
width: Integer dimension
height: Integer dimension
bands: Number of bands in the buffer
format: Band format (all bands must be the same format)
"""
array = array.astype(cls.NUMPY_TYPES[format])
buf = memoryview(array)
image = Image.new_from_memory(buf, width, height, bands, format)
return image
def write_buffer(self, image, resolution):
if VImageAdapter(image).BufferSize() >= self.IMAGE_BUFFER_DISK_THRESHOLD:
logger.debug(
'Buffering resolution {0} to disk'.format(resolution)
)
vipsfile = Image.new_temp_file("%s.v")
tempfile_image = image.write(vipsfile)
return tempfile_image
logger.debug(
'Buffering resolution {0} to memory'.format(resolution)
)
return Image.new_from_memory(
image.write_to_memory(), image.width, image.height, image.bands,
'uchar'
)
def from_numpy_array(cls, array, width, height, bands, format):
"""
Returns a new pyvips.Image created from a NumPy `array` of pixel data.
array: The NumPy array
width: Integer dimension
height: Integer dimension
bands: Number of bands in the buffer
format: Band format (all bands must be the same format)
"""
array = array.astype(cls.NUMPY_TYPES[format])
buf = memoryview(array)
image = Image.new_from_memory(buf, width, height, bands, format)
# Hold on to the numpy array to prevent garbage collection
image._numpy_array = array
return image
def run_on_large_image(root_path: str, original_img: str, ckpt: str,
refine: bool):
'''在一整张图上执行运算
Args:
root_path(str): 从大图中提取的小图所在目录
original_img(str): 大图的路径
ckpt(str): 检查点路径
'''
assert os.path.exists(root_path), '路径不存在: {}'.format(root_path)
assert os.path.exists(original_img), '路径不存在: {}'.format(original_img)
# 配置模型
print('Setting up model.')
model, device = setup(DeepLabV3Res101())
print('Loading model from {}.'.format(ckpt))
model, _ = restore_from(model, ckpt)
model.eval()
file_list = os.listdir(root_path)
original_img_array = vipImage.new_from_file(original_img)
ow, oh = [getattr(original_img_array, key) for key in ['width', 'height']]
large_array = np.zeros((oh, ow), dtype=np.uint8)
# 统计尺度
scales = set()
for file in file_list:
scale, _, _, _ = extract_info_from_filename(file)
scales.add(scale)
# 使用最大尺度的图片作为索引
max_scale = max(scales) if refine else min(scales)
for file in tqdm(file_list, desc='Processing[r={}]'.format(refine)):
scale, x, y, _ = extract_info_from_filename(file)
if scale != max_scale: continue
file_path = os.path.join(root_path, file)
if refine:
refined_image = run_and_refine_single_image(file_path,
ckpt,
False,
model=model,
device=device)
else:
refined_image = run_on_single_image(file_path,
ckpt,
model=model,
device=device)
rh, rw = refined_image.shape # 根据生成图像大小填充最终图像
large_array[x:x + rh, y:y + rw] = refined_image
print('Saving to files...')
# 为输出图片上色，方便查看
colored_array = large_array * 85
# 输出目录
flag = 'refined' if refine else 'norefine'
filename, extension = os.path.basename(original_img).rsplit('.', 1)
output_file = os.path.join(
'./output', '{}_predict_{}_{}.{}'.format(filename, max_scale, flag,
extension))
colored_file = os.path.join(
'./output',
'{}_predict_colored_{}_{}.{}'.format(filename, max_scale, flag,
extension))
vipImage.new_from_memory(large_array.data, ow, oh, 1,
'uchar').write_to_file(output_file)
vipImage.new_from_memory(colored_array.data, ow, oh, 1,
'uchar').write_to_file(colored_file)
print('Large Image File has been saved to {} and {}'.format(
output_file, colored_file))
