1. A Parallel Algorithm for Hardware Implementation of Inverse Halftoning Umair F. Siddiqi 1, Sadiq M. Sait 1 & Aamir A. Farooqui 2 1 Department of Computer Engineering King Fahd University of Petroleum & Minerals, Dhahran 31261, Saudi Arabia 2 Synopsys Inc. Synopsys Module Compiler, Mountain View California, USA

2. Analog halftoning ► The process of rendition of continuous tone pictures on media on which only two levels can be displayed. ► The size of dots are adjusted according to the local print intensity. ► When looked at a distance it gives the impression of the original picture.

3. Digital halftoning ► In digital halftoning the input of the system is a grey-level image having more than two levels for example, 256 levels and the resulting image has only two levels. ► The halftone image is comprised of zeros and ones but gives the impression of the original image from a distance.

4. Inverse halftoning ► Inverse halftoning is the reconstruction of continuous tone picture (e.g. 256 levels) from its halftoned version. ► The input to an inverse halftoning system in an image that consists of zeros and ones and output is an image in which each pixel have value from 256 gray-levels. ► Inverse Halftoning finds application in image compression, printed image processing, scaling, enhancement, etc. ► Inverse halftoning can be for color images but we are concerned with gray-level images and their halftones.

5. Example of Inverse Halftoning Halftone Image Inverse Halftone or grey-level image

6. Demonstration of our Inverse halftoning algorithm ► The next few slides show how inverse halftone operation is performed in our algorithm.

7. Lookup Table (LUT) based Inverse Halftone operation ► The Lookup Table (LUT) method proposed by Mese and Vaidyanathan is used for inverse halftone operation. ► The LUT method uses a template “19pels” to select pixels from the neighborhood of the pixel that is going to be inverse halftone. ► This “19pels” then goes into a LUT which compares the “19pels” with its stored values and returns a gray-level for the input “19pels”.

8. “19pels” Template 12345 678910 111201314 151617 18 The pixel numbered 0 is the one going to be inverse halftoned This pattern is associated with each pixel that is to be inverse halftoned

9. Demonstration of LUT inverse halftoning

10. This is the first “19pels” selected

11. This is the second “19pels” selected

12. This is the third “19pels” selected

13. This is the fourth “19pels” selected

14. Our modification to LUT based Inverse Halftoning

15. Problem of parallel LUT inverse halftone operation ► The LUT method uses one Lookup table that contains inverse halftone values for all “19pels” that are obtained through training set of halftones of standard images. ► To fetch parallel inverse halftone values of more than one 19pels we need to implement multiple copies of the LUT !

16. Our approach to parallel LUT inverse halftoning ► The single large LUT has been divided into many Smaller LUTs (SLUTs). ► Now more than one 19pels can fetch its inverse halftone value from a separate SLUT independent to other parallel 19pels. ► Next problem is to develop a method to send incoming 19pels to separate SLUTs.

17. Method to distinguish 19pels from each other ► The task to send many incoming 19pels to their separate SLUTs is accomplished by defining an operator over 19pels. ► This operator is called Relative XOR Change (RXC). ► When all incoming 19pels are operated through this operator they convert into distinguished values in the range of –t to +t, where t = 19 in our case, but it could be any random integer within a suitable range with respect to total number of SLUTs and hardware complexity.

18. Demonstration of RXC operation

19. RXC Operator for P n 1. P n-1 = “19pels” with the pixel 0 at position (row,col-1); 2. P n = “19pels” with pixel 0 at position (row,col); 3. xor_1= XOR(P n-1, P n ); 4. Magnitude of RXC= |RXC|= Number of Ones(xor_1); 5. Sign of RXC= sgn(RXC)= + when |P n | > |P n-1 | - when |P n | < |P n-1 | - when |P n | < |P n-1 | Note: pixel 0 is the one that is to be inverse halftoned

20. RXC over gray-level halftones I Gray-level 230Corresponding halftone obtained through Floyd and Steinberg Error Diffusion Method

21. RXC over gray-level halftones II Gray-level 130Corresponding halftone obtained through Floyd and Steinberg Error Diffusion Method

22. Magnified look at the halftones I Halftone shows no column-wise periodicity among dots over small 19pels regions Halftone shows column-wise periodicity among dots over small 19pels regions Gray-level 210Gray-level 130

23. Magnified look at the halftones II Halftone shows no periodicity among dots over small 1D 19pels regions Gray-level 120 Halftone shows no periodicity among dots over small 1D 19pels regions Portion of the halftone from image Boat

24. NON Periodic Vibratory RXC Operator ► The operator RXC has been defined that is simple to implement in hardware as well as gives NON periodic vibratory response over most of the gray levels from 0 to 255. ► We have assumed that a gray level image is a composition of many gray levels and obtaining the performance of RXC over individual gray levels can give a clue about its performance on images. ► This assumption is found to be correct in simulation results.

25. Parallel application of RXC

26. Development of parallel table access algorithm with RXC The addition of Slut values from previous pixels simplifies the hardware design

27. Formal Algorithm

28. Simulation ► The algorithm is implemented in MATLAB the performance and quality of inverse halftoning is estimated. ► We assumed LUT inverse halftone operation to be ideal. ► The simulation results show the quality loss with respect to original image that occurred in distribution of parallel “19pels” to different SLUTs through RXC. ► This pixel loss is compensated through replicating gray level values from the neighbors.

29. Sample Image I PSNR= 34.7880 peppers

30. Sample Image II PSNR= 32.5685 lena

31. Sample Image III PSNR= 28.1264 mandrill

32. Hardware Implementation ► This section shows the hardware implementation of the proposed parallel algorithm in terms of block diagrams. ► The specification of the hardware design is: 1. Parallel Pixels to be inverse halftone= n= 15 2. Number of SLUTs= 19

33. Two Blocks of hardware Implementation ► The hardware system can be divided into two blocks: 1. RXC and modulus operators 2. 19pels to gray-level decoders

34. RXC and modulus operators ► RXC and modulus operators components are responsible for the following tasks: Input: 19pels Output: SLUT numbers Slut 1. Accept 19pels from the halftone image and assign a sequence number to each entered 19pels. 2. Perform RXC operation on all 19pels. 3. Add the Slut value of the 19pels that has preceding sequence number to the current result. 4. Then take mod of the current result with a fixed number i.e. 19 in our case to obtain Slut value for the current 19pels. 5. The above three steps are pipelined so new 19pels are coming in while the current 19pels are in process.

35. RXC and modulus Block Diagram RXC calculation for 19pels P n P n-1 and P n are two 19pels among all 19pels to be inverse halftoned in parallel. Slut is the Smaller LUT number where the concerned 19pels should go to fetch its inverse halftone value.

36. Hardware Design of RXC and modulus Operator ► The next slides can show the hardware design of RXC operator for a 19pels pattern named P n with the following parameters: ► Parallel pixels to be inverse halftoned at a time= 15 ► Total number of SLUTs= 19, therefore, Slut is from 0 to 19.

37. Determination of Slut from RXC

38. Block diagram showing gray-level decoding process

39. Routing of a 19pels to 5 th SLUT

40. Routing of a 19pels to 16 th SLUT

41. Routing of a 19pels to 3 rd SLUT

42. Routing of a 19pels to 17 th SLUT

43. SLUT i(i=16)

44. Quality of inverse halftones Image Halftone Algorithm %pixel coverage w/o pixel compensation PSNR with pixel compensation Boat FS ED 65.086430.3749 Clock 70.666730.1671 Peppers 68.943328.5484 Boat GN ED 63.753128.7139 Clock 69.876531.2554 Peppers 68.950929.0077 Boat EG ED 67.308632.1370 Clock 68.592629.9289 Peppers 69.990528.5483

45. Comparison to halftone 256*256 Algorithm in [7] Proposed Algorithm Cycles/pixel 1 0.066 LUT size 5.1 K entries 19 K entries Latency 4 clock cycles 17 clock cycles Time taken 691.3502 ms 45.6389 ms

46. Conclusion and Future Work ► A parallel implementation for inverse halftone has been presented. ► Results can be improved by improving the operators and training. ► Results obtained are encouraging.

47. Method to generate contents of SLUT ► The algorithm is applied on images in a training set and Sluts values are obtained. ► The 19pels then placed in the SLUT given by the corresponding Slut value.

48. Properties of SLUTs ► The SLUTs were developed using training set composed of FS ED halftone images of Boat and Peppers of size 256x256-pixels. ► The size of one SLUT is found to be 2.5K entries. ► The summation of entries in all 19 SLUTs comes to be 42.6K. ► The size of LUT in single LUT method is 9.86K entries, however, if the single LUT method is implemented multiple times for 15 parallel pixels the total size could become 148K entries. ► In this way, our method can provide 3.5 times decrease in lookup table size over single LUT based method.

49. Behavior of RXC over Grey-level halftones NON Periodic Vibratory ResponsePeriodic Vibratory Response Gray level 130Gray level 210 Halftones obtained through Floyd & Steinberg Error Diffusion Method

50. Representation of RXC values on number line Periodic Vibratory Values RXC values to be used in SLUT access are calculated by adding the RXC to the RXC of the previous “19pels” That is: RXC for SLUT of P n (Slut)= RXC of P n-1 + RXC of P n-2 (n) From the number line we can see that adding RXC over previous values gives zero or constant result, therefore, we need NOT periodic vibratory response from RXC operator.

51. Modified RXC I ► At present, RXC is a comparative operator that it gives values in comparison to the previous 19pels. ► This behavior of RXC can give different Slut values if some 19pels are replaced from the image. ► Therefore, we are required to store same 19pels in more than one table.

52. Modified RXC II ► Let us define a standard value for RXC= The value of 19pels at which the histogram of 19pels present in training set images can be divided into two portions. The value of 19pels at which the histogram of 19pels present in training set images can be divided into two portions. ► We find Slut value of each 19pels with respect to this standard value. ► That way we can have almost uniform table size with no repetition of same value in different tables.

53. Example histogram The mean 19pels= 0111110100011100110