Paddlepaddle 28 build an automatic coder based on convolution

Automatic coder can be used for image compression 、 Image denoising and other tasks , It is composed of encoder and decoder , Its essence is to encode and decode the data . The output of the encoder is used for the input of the decoder , In automatic coders, it is generally considered that the function of the decoder is the inverse transformation of the encoder . If the encoding and decoding process is almost lossless , That is image compression ; If the process of compression and decoding is lossy 【 Not really detrimental , It's different 】（ such as ： The decoded image is clearer than the original image , Image denoising 、 Defogging 、 Go to the rain ; The resolution is improved , Image super-resolution reconstruction ）.

When decoding lossless , That is to say, image compression , The input and output of the automatic coder are the same .

When decoding is damaged , That is to say, image optimization , The input and output of the automatic coder are different , The difference between input and output is the experience learned by the model .

stay paddle Within the framework of , It's hard to find an example of an automatic coder . therefore , Bloggers have made a simple implementation , And minist Simply test the data .

1、 The core structure of automatic coding machine

In the usual automatic coder , The structure of encoder and decoder is symmetrical , The encoder is responsible for extracting features （ For automatic coder based on convolution, feature extraction is mainly through conv With down sampling ）, The decoder is responsible for restoring features （ For automatic decoding machine based on convolution, feature extraction is mainly through conv With up sampling ）.

The encoder is composed of convolution pooling , Convolution feature extraction , Pool brushing characteristics ; The decoder consists of up sampling and convolution , There are two ways to realize up sampling, interpolation and transpose convolution . Interpolation method refers to UpSample The way , This method is derived from the traditional image scaling method ; Transpose convolution is Conv2DTranspose, Is another convolution implementation , Through to feature map fill 0 Then do convolution to realize sampling .Conv2DTranspose For more details, please refer to the following links . It should be noted that UpSample Is parameterless , Without training ; and Conv2DTranspose