Reducing image resolution in deep convolutional networks

1. A method of reducing image resolution in a deep convolutional network (DCN), comprising:

dynamically selecting, for each layer of a plurality of layers of the DCN, a reduction factor to be applied to an input image; and

adjusting the DCN based at least in part on the reduction factor selected for each layer.

2. The method of claim 1 , in which dynamically selecting the reduction factor comprises:

randomly selecting the reduction factor for each layer; and

adjusting each reduction factor based at least in part on performance loss and a level of complexity of the DCN.

3. The method of claim 1 , in which dynamically selecting the reduction factor comprises:

selecting the reduction factor, for each layer, based at least in part on computational complexity of each layer of the DCN and/or a fraction of energy in high frequency components of the input image at a given layer; and

adjusting each reduction factor based at least in part on performance loss and a level of complexity of the DCN.

4. The method of claim 1 , in which adjusting the DCN further comprises adjusting weights in each layer of the DCN based at least in part on the reduction factor.

5. The method of claim 1 , in which adjusting the DCN further comprises trimming synaptic connections in pooling layers of the DCN based at least in part on the selected reduction factor.

6. The method of claim 1 , in which adjusting the DCN further comprises fine-tuning the DCN.

7. The method of claim 6 , in which the fine-tuning comprises fine-tuning selected layers while maintaining constant weights in at least one other layer.

8. The method of claim 7 , in which the selected layers comprise convolutional layers and the least one other layer comprises a fully-connected layer.

9. The method of claim 1 , further comprising:

storing one fully-connected layer for both an original DCN and the adjusted DCN; and

storing a first convolutional layer for the original DCN and a second convolutional layer for the adjusted DCN.

10. An apparatus for reducing image resolution in a deep convolutional network (DCN), comprising:

means for dynamically selecting, for each layer of a plurality of layers of the DCN, a reduction factor to be applied to an input image; and

means for adjusting the DCN based at least in part on the reduction factor selected for each layer.

11. The apparatus of claim 10 , in which the means for dynamically selecting the reduction factor comprises:

means for randomly selecting the reduction factor for each layer; and

means for adjusting each reduction factor based at least in part on performance loss and a level of complexity of the DCN.

12. The apparatus of claim 10 , in which the means for dynamically selecting the reduction factor comprises:

means for selecting the reduction factor, for each layer, based at least in part on computational complexity of each layer of the DCN and/or a fraction of energy in high frequency components of the input image at a given layer; and

means for adjusting each reduction factor based at least in part on performance loss and a level of complexity of the DCN.

13. The apparatus of claim 10 , in which the means for adjusting the DCN further comprises means for adjusting weights in each layer of the DCN based at least in part on the reduction factor.

14. The apparatus of claim 10 , in which the means for adjusting the DCN further comprises means for trimming synaptic connections in pooling layers of the DCN based at least in part on the selected reduction factor.

15. The apparatus of claim 10 , in which the means for adjusting the DCN further comprises means for fine-tuning the DCN.

16. The apparatus of claim 15 , in which the means for fine-tuning comprises means for fine-tuning selected layers while maintaining constant weights in at least one other layer.

17. The apparatus of claim 16 , in which the selected layers comprise convolutional layers and the least one other layer comprises a fully-connected layer.

18. The apparatus of claim 10 , further comprising:

means for storing one fully-connected layer for both an original DCN and the adjusted DCN; and

means for storing a first convolutional layer for the original DCN and a second convolutional layer for the adjusted DCN.

19. An apparatus for reducing image resolution in a deep convolutional network (DCN), comprising:

a memory unit; and

at least one processor coupled to the memory unit, the at least one processor configured:

to dynamically select, for each layer of a plurality of layers of the DCN, a reduction factor to be applied to an input image; and

to adjust the DCN based at least in part on the reduction factor selected for each layer.

20. The apparatus of claim 19 , in which the at least one processor is further configured:

to randomly select the reduction factor for each layer; and

to adjust each reduction factor based at least in part on performance loss and a level of complexity of the DCN.

21. The apparatus of claim 19 , in which the at least one processor is further configured:

to select the reduction factor, for each layer, based at least in part on computational complexity of each layer of the DCN and/or a fraction of energy in high frequency components of the input image at a given layer; and

to adjust each reduction factor based at least in part on performance loss and a level of complexity of the DCN.

22. The apparatus of claim 19 , in which the at least one processor is further configured to adjust weights in each layer of the DCN based at least in part on the reduction factor.

23. The apparatus of claim 19 , in which the at least one processor is further configured to trim synaptic connections in pooling layers of the DCN based at least in part on the selected reduction factor.

24. The apparatus of claim 19 , in which the at least one processor is further configured to fine-tune the DCN.

25. The apparatus of claim 24 , in which the at least one processor is further configured to fine-tune selected layers while maintaining constant weights in at least one other layer.

26. The apparatus of claim 25 , in which the selected layers comprise convolutional layers and the least one other layer comprises a fully-connected layer.

27. The apparatus of claim 19 , in which the at least one processor is further configured:

to store one fully-connected layer for both an original DCN and the adjusted DCN; and

to store a first convolutional layer for the original DCN and a second convolutional layer for the adjusted DCN.

28. A non-transitory computer-readable medium for reducing image resolution in a deep convolutional network (DCN), the non-transitory computer-readable medium having program code recorded thereon, the program code being executed by a processor and comprising:

program code to dynamically select, for each layer of a plurality of layers of the DCN, a reduction factor to be applied to an input image; and

program code to adjust the DCN based at least in part on the reduction factor selected for each layer.

29. The computer-readable medium of claim 28 , further comprising:

program code to randomly select the reduction factor for each layer; and

program code to adjust each reduction factor based at least in part on performance loss and a level of complexity of the DCN.

30. The computer-readable medium of claim 28 , further comprising:

program code to select the reduction factor, for each layer, based at least in part on computational complexity of each layer of the DCN and/or a fraction of energy in high frequency components of the input image at a given layer; and

program code to adjust each reduction factor based at least in part on performance loss and a level of complexity of the DCN.

31. The computer-readable medium of claim 28 , further comprising program code to adjust weights in each layer of the DCN based at least in part on the reduction factor.

32. The computer-readable medium of claim 28 , further comprising program code to trim synaptic connections in pooling layers of the DCN based at least in part on the selected reduction factor.

33. The computer-readable medium of claim 28 , further comprising program code to fine-tune the DCN.

34. The computer-readable medium of claim 33 , further comprising program code to fine-tune selected layers while maintaining constant weights in at least one other layer.

35. The computer-readable medium of claim 34 , in which the selected layers comprise convolutional layers and the least one other layer comprises a fully-connected layer.

36. The computer-readable medium of claim 28 , further comprising:

program code to store one fully-connected layer for both an original DCN and the adjusted DCN; and

program code to store a first convolutional layer for the original DCN and a second convolutional layer for the adjusted DCN.