Question

我正在尝试保存带有显示边界框的图像，只是为了测试我的注释文件是否正常工作。

一切都很好：图像写入磁盘，边界框位于正确的位置，依此类推。除了所有颜色都是倒置的。所以它看起来像原始图像的负片。

这是我的代码：

```

import tensorflow as tf
import numpy as np
from PIL import Image

def read_processed_data(filename, num_show):
    """ Reads in the processed data file and displays the
        given number of images, along with the bounding boxes.
    """
    with open(filename, 'r') as f:
        i = 0

        while i < num_show:
            for line in f:
                filename = line.rstrip()
                next_line = f.readline()
                num_faces = int(next_line.rstrip())
                face_num = 0

                #while face_num < num_faces:
                bb_line = f.readline().rstrip()
                y1, x1, y2, x2 = bb_line.split(',')
                y1 = float(y1)
                x1 = float(x1)
                y2 = float(y2)
                x2 = float(x2)

                box = [y1, x1, y2, x2]

                return box, filename


with tf.Session() as sess:
    bb, fn = read_processed_data("processed.txt", 1)
    image = tf.image.decode_image(tf.read_file(fn))

    image_as_float = tf.cast(image, dtype = tf.float32)
    image_4d = tf.expand_dims(image_as_float, 0)

    bb_2d = tf.expand_dims(bb, 0)
    bb_3d = tf.expand_dims(bb_2d, 0) # Box has to be 3d for the drawing to work
    bb_image = tf.image.draw_bounding_boxes(image_4d, bb_3d)
    bb_image_uint = tf.image.convert_image_dtype(bb_image, dtype = tf.uint8)
    bb_image_uint_3d = tf.reshape(bb_image_uint, [940, 650, 3]) # Reduce rank from 4 to 3
    data = bb_image_uint_3d.eval()

    base_fn = fn.split('.')[0]
    Image.fromarray(data).save(base_fn + "_bb.jpg")

```

我搜索过tensorflow文档无济于事。我还尝试了np.roll()和来自PIL rotate image colors (BGR -> RGB)的其他建议，但没有运气;这些方法能够改变颜色，但不能改变正确的颜色。

https://imgur.com/a/cclKJ显示顶部的原始图像（没有边界框），以及下面的结果图像（带有颜色问题，以及边界框）。

Answer 1

我使用saturate_cast()解决了这个问题。看来uint8 - ＆gt; float32或float32 - ＆gt; uint8转换（我怀疑后者）导致溢出。

代码的固定部分是

image_as_float = tf.saturate_cast(image, dtype = tf.float32)
image_4d = tf.expand_dims(image_as_float, 0)    # Add in a batch dimension (of size 1)

bb_2d = tf.expand_dims(bb, 0)   # Add in dimension of size 1 (num_bounding_boxes)
bb_3d = tf.expand_dims(bb_2d, 0)    # Add in dimension of size 1 (batch)
bb_image = tf.image.draw_bounding_boxes(image_4d, bb_3d)
bb_image_uint = tf.saturate_cast(bb_image, dtype = tf.uint8)

感谢de1提出的建议，就像我在测试固定代码一样。

使用TensorFlow和PIL写入图像时的反色

1 个答案: