2018-08-13 image data processng

pretreatment

TFRecord

input data format
Example{Features{map<String,Feature{BytesList or FloatList or Int64List}>}}
write

• step1: setup a writer with filename for TFRecord
• step2: get the images data ，labels data ，and etc.
• step3: piecewise transformation{
  setup a example with image-concerned data
  write its serialized sequence into writer
  }
  read
• step1: setup a reader with filepath
• step2: read the serialized example
• step3: parse_single_example(serialized_example,features={'': tf.FixedLenFeature([],tf.dtype,...)})
• step4: images= tf.decode_raw(features['image_raw'], tf.uint8)
  other cast(features,dtype)
• step5: multi threads for processing

image processing functions

• encoding
  objects: jpeg png

image_raw_data=tf.gfile.FastGFile('image_path','r').read()
with tf.session() as sess:
       img_data=tf.image.decode_jpeg(image_raw_data)
       print(img_data.eval())
       plt.imshow(img_data.eval())
       plt.show()
       img_data=tf.image_convert_iamge_dtype(img_data,dtype=tf.float32)
#encode 

• scale adjustment
  tf.image.resize_images(img_data, 300,300,method=0)
  method:

  method value	algorithms
  0	Bilinear interpolation
  1	Nearest neighbor interpolation
  2	Bicubic interpolation
  3	Area interpolation

tf.image.resize_image_with_crop_or_pad(img_data, len_scale,wid_scale)
tf.image.central_crop(img_data,proportion)
tf.image.flip_up_down(...)#left_right
tf.image.transpose_image(image_data)
tf.image.random_flip_...

• color
  tf.image.adjust_brightness(img_data,-0.5)
  tf.image.random_brightness(img_data,max_delta)
  tf.image.adjust_contrast(img_data,-5)
  tf.image.random_contrast(...,lower,upper)
  tf.image.adjust_hue(image_data,)#0.1,0.3,0.6,0.9
  tf.image.random_hue(,max_delta)
  tf.image.adjust_saturation(...,...,...)
  tf.image.per_image_whitening(...)#standardize the image
• marking frame
  tf.image.draw_bounding_boxes(img_data,coordinates of boxes)