"Image Processing" TensorFlow: Simple super-resolution reconstruction and pit

Super-resolution reconstruction is a hot spot in the field of image restoration, which can minimize the signal of original scene in the case of limited hardware, and plays an important role in the fields of astronomical exploration and microscopic imaging. Imaging equipment for the object imaging, because the distance, imaging will be blurred, can be analogous to multi-scale Gaussian filter, limited by imaging functions, imaging pixels can not achieve the most ideal conditions, the analogy to the original image for a sample. Super-resolution reconstruction is to restore the original image under this condition.
Suppose God has the best imaging equipment, imaging X; Our imaging equipment is B, Gaussian filter template is set to G; In order to prevent the problem of morbid, add lasso regular. So there are: Argmin [subsampling (CONV (x,g)) −b]2+λx argmin\ [subsampling (CONV)) X,g X Now the problem is, TensorFlow how to represent subsampling and optimize it.
TensorFlow supports the following image scaling/sampling:

Tf.image.resize_images, support nearest neighbor, Bilinear, double three times scaling method Tf.nn.max_pool Maximum value sampling

Tf.nn.avg_pool sampling under Mean value

Now let's test it one by one. The image is sampled after three times times:
　　
1, Tf.image.resize_images, bilinear sampling, ringing is not serious, a lot of stripes:
　　
2, Tf.nn.max_pool, no stripes, ringing, but there is a lot of noise, parameters tuned several times there is no better effect:
　　
3, Tf.nn.avg_pool, no stripes, noise, there is a ringing, and the original image of the color darkened, contrast decline:
　　
4, to make a comparison with the original artwork
　　

Can see, the best effect is avg_pool, in only Gaussian template parameters, completely no other prior information, a second to get this result, has been very surprising. Guess Image-resize and Max_pool in fact in the upper and lower samples are missing quite a lot of signals, and avg_pool retain the most signals, so the reconstruction effect is better.
The Fast-neural-style article mentions the use of perceptual features for super-resolution reconstruction of images, which can be reconstructed with the same style of detail, which requires a large number of images to be trained using a build network, or VGG slowly compute perceptual features to mimic style details.