Processing images using hybrid infinite impulse response (TTR) and finite impulse response (FIR) convolution block

1. An apparatus for processing images, the apparatus comprising:

at least one memory;

instructions; and

processor circuitry to execute the instructions to implement:

a hybrid infinite impulse response-finite impulse response (IIR-FIR) convolution to process an image to generate processed image information, the hybrid IIR-FIR convolution including a vertical infinite impulse response (IIR) component to approximate a vertical convolution; and

at least one of a channel compressor or a differential pulse-code modulation (DPCM) encoder, the channel compressor to reduce a number of bits to be stored in a skip line buffer, the DPCM encoder to compress skip lines in a neural network.

2. The apparatus of claim 1 , wherein the hybrid IIR-FIR convolution includes a horizontal finite impulse response (FIR) convolution.

3. The apparatus of claim 1 , wherein the hybrid IIR-FIR convolution includes a pointwise FIR convolution.

4. The apparatus of claim 1 , wherein the vertical IIR component is implemented by a spatially-recurrent neural network.

5. The apparatus of claim 1 , wherein the processor circuitry is to implement a max pooling layer to perform a downscaling of feature maps by a factory of four.

6. The apparatus of claim 1 , wherein the processor circuitry is to implement a trainable vision scaler to downscale the image in response to detecting that a resolution of the image exceeds a threshold.

7. The apparatus of claim 1 , wherein the apparatus is to receive six channels of image information and output up to six channels of processed image information.

8. The apparatus of claim 1 , wherein the apparatus includes a fully pipelined inline hardware device trainable to perform a plurality of image processing tasks.

9. A method for processing images, the method comprising:

processing an image with a hybrid infinite impulse response-finite impulse response (IIR-FIR) convolution to generate processed image information, the hybrid IIR-FIR convolution including an approximated vertical convolution;

at least one of encoding a feature map or compressing the feature map, the feature map included in the processed image information, the encoding of the feature map using differential pulse-code modulation (DPCM) to compress skip lines in a neural network, the compressing of the feature map to reduce a number of bits to be stored in a skip line buffer; and

outputting, with a processor, the processed image information.

10. The method of claim 9 , further including downscaling the image with a trainable vision scaler and upsampling the processed image information with a micro upscaler in response to detecting that a resolution of the image exceeds a predetermined threshold.

11. The method of claim 9 , further including performing a downscaling of the feature map by a factory of four.

12. The method of claim 9 , wherein the processing of the image includes denoising the image.

13. The method of claim 9 , wherein the processing of the image includes deblurring the image.

14. The method of claim 9 , wherein the processing of the image includes missing color information from context in the image.

15. The method of claim 9 , wherein the processing of the image includes demosaicing the image.

16. The method of claim 9 , wherein the processing of the image includes spatially varying exposure levels in the image.

17. An article of manufacture including at least one memory device, the article of manufacture comprising instructions that, when executed, cause a computing device to at least:

process an image with an approximated vertical convolution to generate processed image information;

at least one of (i) encode a feature map using differential pulse-code modulation (DPCM) to compress skip lines in a neural network, or (ii) compress the feature map to reduce a number of bits to be stored in a skip line buffer, the feature map included in the processed image information; and

output the processed image information.

18. The article of manufacture of claim 17 , wherein the instructions cause the computing device to:

downscale the image in response to detecting that a resolution of the image exceeds a predetermined threshold;

process the downscaled image to generate the processed image information; and

upsample the processed image information.

19. The article of manufacture of claim 17 , wherein the instructions cause the computing device to perform a downscaling of feature maps by a factory of four.