tf.image.resize()

作用

     ~~~~     Resize the image using the specified method.
     ~~~~     If the original aspect ratio is different from the size,The resized image will be distorted. to avoid distortion,可用tf.image.resize_with_pad（）

参数

tf.image.resize(
    images,                         # 形状 [batch, height, width, channels] 的 4-D tensor or shape [height, width, channels] 的 3-D 张量.
    size,                           # 2 个元素的一维 int32 张量：new_height,new_width. 图像的新尺寸.
    method=ResizeMethod.BILINEAR,   # 一个 image.ResizeMethod,or the equivalent string. The default is bilinear.
    preserve_aspect_ratio=False,    # Whether to preserve the aspect ratio.
                                    # true,The image will be resized to fit the size,While preserving the aspect ratio of the original image. 
                                    # If the size is larger than the current size of the image,则放大图像. 默认为false.
    antialias=False,                # Whether to use when downsampling the imageanti-aliasing filter.
    name=None                       # 此操作的名称（可选）.
)

     ~~~~     其中,Image.ResizeMethod()的选项有：
     ~~~~     • bilinear： 双线性插值.若antialisa为true,Then it becomes the radius when downsampling1的hat/tent filter.
     ~~~~     • lanczos3： Lanczos kernel with radius 3.High quality practical filter,But there may be someringing,Especially on composite images.
     ~~~~     • lanzos5：Lanczos kernel with radius 5.Very high quality filter,But there may be stronger onesringing.
     ~~~~     • bicubic：三次插值.相当于Catmull-Rom kernel,与lanczos3kernel相比较,质量相当好,速度更快,Especially when upsampling.
     ~~~~     • gaussian：Gaussian kernel with radius 3, sigma = 1.5 / 3.0.
     ~~~~     • nearest：最近邻插值. When used with nearest neighbor interpolation,antialias无效.
     ~~~~     • area：Antialiased resampling using region interpolation. When used with area interpolation,antialias没有效果.
     ~~~~     •mitchellcubic：Mitchell-Netravali Cubic non-interpolating filter. For composite images（Especially those images that lack proper pre-filtering）,ringing比Cubic interpolant of Keys少,Not as sharp.

返回值

     ~~~~     If images was 4-D, a 4-D float Tensor of shape [batch, new_height, new_width, channels].
     ~~~~     If images was 3-D, a 3-D float Tensor of shape [new_height, new_width, channels].

例子

import tensorflow as tf                               # 导入tensorflow
img_path ='C:\\Users\\xxx\\.keras\\datasets\\flower_photos\\roses\\6158504080_b844a9ae05.jpg'    # 输入图片路径
img_raw = tf.io.read_file(img_path)                   # 读取图片内容,返回值时string的tensor
img_tensor = tf.image.decode_image(img_raw)           # 对输入的string的tensor进行解码
print(img_tensor.shape)                               # Enter the decoded image shape
print(img_tensor.dtype)                               # Enter the decoded image representation type
img_final = tf.image.resize(img_tensor, [192, 192])   # 调整图像大小
print(img_final.shape)                                # Enter the resized image size

(333, 500, 3)
<dtype: 'uint8'>
(192, 192, 3)

tf.image.resize_with_pad()

作用

     ~~~~     Resize and fill the image to the target width and height.
     ~~~~     No distortion by keeping the aspect ratio the same,Resize the image to the target width and height. If the target size does not match the image size,will resize the image,It is then padded with zeros to match the requested dimensions.

参数

tf.image.resize_with_pad(
    image,                              # 形状 [batch, height, width, channels] 的 4-D tensor or shape [height, width, channels] 的 3-D 张量.
    target_height,                      # 目标高度
    target_width,                       # 目标宽度
    method=ResizeMethod.BILINEAR,       # The method used to resize the image. 见 image.resize()
    antialias=False                     # Whether to use antialiasing when resizing. 参见“image.resize()”.
)

返回值

     ~~~~     Resized and padded image.
     ~~~~     If images was 4-D, a 4-D float Tensor of shape [batch, new_height, new_width, channels].
     ~~~~     If images was 3-D, a 3-D float Tensor of shape [new_height, new_width, channels].

主要参考：
tf.image.resize | TensorFlow Core v2.9.1 (google.cn)
tf.image.resize_with_pad | TensorFlow Core v2.9.1 (google.cn)