Bilateral Filter

2-D bilateral filtering

• 

Libraries:
Vision HDL Toolbox / Filtering

Description

The Bilateral Filter block filters images while preserving edges. Some applications of bilateral filtering are denoising while preserving edges, separating texture from illumination, and cartooning to enhance edges. The filter replaces each pixel at the center of a neighborhood by an average that is calculated using spatial and intensity Gaussian filters. The block determines the filter coefficients from:

• Spatial location in the neighborhood (similar to a Gaussian blur filter)

• Intensity difference from the neighborhood center value

The block provides two standard deviation parameters for independent control of the spatial and intensity coefficients.

Examples

Generate Cartoon Images Using Bilateral Filtering

Generate cartoon lines and overlay them onto an image.

Open Script

Ports

This block uses a streaming pixel interface with a pixelcontrol bus for frame control signals. This interface enables the block to operate independently of image size and format. All Vision HDL Toolbox™ blocks use the same streaming interface. The block accepts and returns a scalar pixel value and a bus that contains five control signals. The control signals indicate the validity of each pixel and its location in the frame. To convert a frame (pixel matrix) into a serial pixel stream and control signals, use the Frame To Pixels block. For a full description of the interface, see Streaming Pixel Interface.

Input

pixel — Input pixel stream

scalar | vector

This block supports single pixel streaming or multipixel streaming. For single pixel streaming, specify a single input pixel as a scalar intensity value. For multipixel streaming, specify a vector of two, four, or eight pixel intensity values. For details of how to set up your model for multipixel streaming, see Filter Multipixel Video Streams.

This block does not support multicomponent streaming. To process multicomponent streams, replicate the block for each component. The pixelcontrol bus for all components is identical, so you can connect a single bus to multiple replicated blocks.

The software supports double and single data types for simulation, but not for HDL code generation.

Data Types: uint | int | fixed point | double | single

ctrl — Control signals associated with pixel stream

pixelcontrol bus

The pixelcontrol bus contains five signals. The signals describe the validity of the pixel and its location in the frame. For more information, see Pixel Control Bus.

For multipixel streaming, each vector of pixel values has one set of control signals. Because the vector has only one valid signal, the pixels in the vector must be either all valid or all invalid. The hStart and vStart signals apply to the pixel with the lowest index in the vector. The hEnd and vEnd signals apply to the pixel with the highest index in the vector.

Data Types: bus

Output

pixel — Output pixel stream

scalar | vector

Output pixel stream, returned as a scalar value representing intensity, or as a vector of two, four, or eight pixel intensity values. The dimensions and data type of the output pixel port match the dimensions and data type of the input pixel port.

The software supports double and single data types for simulation, but not for HDL code generation.

Data Types: uint | int | fixed point | double | single

ctrl — Control signals associated with pixel stream

pixelcontrol bus

The pixelcontrol bus contains five signals. The signals describe the validity of the pixel and its location in the frame. For more information, see Pixel Control Bus.

For multipixel streaming, each vector of pixel values has one set of control signals. Because the vector has only one valid signal, the pixels in the vector must be either all valid or all invalid. The hStart and vStart signals apply to the pixel with the lowest index in the vector. The hEnd and vEnd signals apply to the pixel with the highest index in the vector.

Data Types: bus

Parameters

Main

Neighborhood size — Size of image region to average

3×3 (default) | 5×5 | 7×7 | 9×9 | 11×11 | 13×13 | 15×15

Size of the image region used to compute the average, specified as an N-by-N pixel square.

Spatial standard deviation — Spatial standard deviation target

0.5 (default) | positive real number

Spatial standard deviation target used to compute coefficients for the spatial Gaussian filter, specified as a positive real number. This parameter has no limits, but recommended values are from 0.1 to 10. At the high end, the distribution becomes flat and the coefficients are small. At the low end, the distribution peaks in the center and has small coefficients in the rest of the neighborhood. These boundary values also depend on the neighborhood size and the data type used for the coefficients.

Intensity standard deviation — Intensity standard deviation target

0.5 (default) | positive real number

Intensity standard deviation target used to compute coefficients for the intensity Gaussian filter, specified as a positive real number. This parameter has no limits, but recommended values are from 0.1 to 10. At the high end, the distribution becomes flat and the coefficients are small. At the low end, the distribution peaks in the center and has small coefficients in the rest of the neighborhood. These boundary values also depend on the neighborhood size and the data type used for the coefficients.

When the intensity standard deviation is large, the bilateral filter acts more like a Gaussian blur filter, because the intensity Gaussian has a lower peak. Conversely, when the intensity standard deviation is smaller, edges in the intensity are preserved or enhanced.

Padding method — Method for padding boundary of input image

Constant (default) | Replicate | Symmetric | Reflection | None

Select one of these methods for padding the boundary of the input image. For more information about these methods, see Edge Padding.

• Constant — Interpret pixels outside the image frame as having a constant value.

• Replicate — Repeat the value of pixels at the edge of the image.

• Symmetric — Set the value of the padding pixels to mirror the edge of the image.

• Reflection — Set the value of the padding pixels to reflect around the pixel at the edge of the image.

• None — Exclude padding logic. The block does not set the pixels outside the image frame to any particular value. This option reduces the hardware resources used by the block and the blanking required between frames but affects the accuracy of the output pixels at the edges of the frame. To maintain pixel stream timing, the output frame is the same size as the input frame. However, to avoid using pixels calculated from undefined padding values, mask off the KernelSize/2 pixels around the edge of the frame for downstream operations. For details, see Increase Throughput by Omitting Padding.

Padding value — Value used to pad boundary of input image

0 (default) | integer

Specify an integer to pad the boundary of the input image. The block casts this value to the same data type as the input pixel.

Dependencies

To enable this parameter, set the Padding method parameter to Constant.

Line buffer size — Size of line memory buffer

2048 (default) | positive integer

Size of the line memory buffer, specified as a positive integer. Choose a power of two that accommodates the number of active pixels in a horizontal line. If you specify a value that is not a power of two, the buffer uses the next largest power of two.

When you use multipixel input, this value must accommodate (Active pixels per line)/(Number of pixels) + 2.

Data Types

Rounding mode — Rounding method for internal fixed-point calculations

Floor (default) | Ceiling | Convergent | Nearest | Round | Zero

Specify a rounding method for internal fixed-point calculations.

Saturate on integer overflow — Overflow mode for internal fixed-point calculations

on (default) | off

When the input is any integer or fixed-point data type, the algorithm uses fixed-point arithmetic for internal calculations. By default, fixed-point values saturate on overflow. This option does not apply when the input data type is single or double.

Coefficients — Method to determine data type of filter coefficients

Inherit: Same as first input (default) | fixdt(1,16,0) | data type expression

Specify an unsigned data type that can represent values less than 1. The coefficients usually require a data type with more precision than the input data type. The block calculates the coefficients based on the neighborhood size and the values of Intensity standard deviation and Spatial standard deviation. Larger neighborhoods spread the Gaussian function such that each coefficient value is smaller. A larger standard deviation flattens the Gaussian so that the coefficients are more uniform in nature, and a smaller standard deviation produces a peaked response.

Note

If you try a data type and after quantization, more than half of the coefficients become zero, the block issues a warning. If all the coefficients are zero after quantization, the block issues an error. These messages mean that the block was unable to express the requested filter by using the data type specified. To avoid this issue, choose a higher-precision coefficient data type or adjust the standard deviation targets.

Output — Method to determine data type of output pixels

Inherit: Same as first input (default) | fixdt(1,16,0) | data type expression

The filtered pixel values are cast to this data type.

Tips

• When you use a block with an internal line buffer inside an Enabled Subsystem (Simulink), the enable signal pattern must maintain the timing of the pixel stream, including the minimum blanking intervals. If the enable pattern corrupts the timing of the pixel stream, you might see partial output frames, corrupted pixel stream control signals, or mismatches between Simulink^® and HDL simulation results. You may need to extend the blanking intervals to accommodate for cycles when the enable is low. For more information, see Configure Blanking Intervals.

Algorithms

The bilateral filter can be described as a Gaussian filter in the spatial dimension that modifies the coefficients of a second Gaussian filter that operates on intensity.

The algorithm stores N-1 lines so that it can form an N-by-N matrix of pixels matching the Neighborhood size. Then it applies two Gaussian filters on each neighborhood. The filter coefficients are calculated from the spatial and intensity standard deviations.

The Subtract Center operation produces a pixel value of zero at the center of the neighborhood. For hardware implementation, and for simulation of fixed-point or integer data types, the calculation in the dashed region is implemented with a lookup table of precomputed values for each pixel. Because the center value is always zero, u² and e^u are always one and are not computed. For floating-point input, the simulation computes u² and e^u as shown. The output of the dashed region uses the coefficient data type that you specified. The Q blocks in the diagram show quantization points.

The algorithm implements the final normalization step with a reciprocal lookup table in the hardware implementation. The lookup table has 2048 locations, so the coefficient sum is quantized to the most significant 11 bits. The reciprocal values use the output data type that you specified, plus a minimum of two integer bits if the data type does not already include them. The reciprocal lookup value for a zero sum is the maximum representable value in the coefficient data type. For floating-point normalization, the simulation detects a zero sum and instead divides by eps() of the dividend.

The output pixel value is then cast to the output data type that you specified. The filter uses the entire range of the data type, so if your color space uses less than the full range, you may need to rescale the pixel values.

Note

When filtering multicomponent (color) pixels, false colors can occur, unless the operation is done in a color space based on human perception, such as CIELab. Bilateral filtering of the R'G'B' color space is not recommended.

Latency

The latency of the block is the line buffer latency plus the latency of the kernel calculation. The line buffer latency includes edge padding by default. The latency of the padding operation depends on the size of the kernel. If edge padding is not necessary for your design, you can reduce the latency by setting the Padding method parameter to None. When you use this option, the block latency does not depend on your kernel size. To determine the exact latency for any configuration of the block, measure the number of time steps between the input and output control signals.

Note

When you use edge padding, use a horizontal blanking interval of at least twice the kernel width. This interval lets the block finish processing one line before it starts processing the next one, including adding padding pixels before and after the active pixels in the line.

The horizontal blanking interval is equal to TotalPixelsPerLine – ActivePixelsPerLine or, equivalently, FrontPorch + BackPorch. Standard streaming video formats use a horizontal blanking interval of about 25% of the frame width. This interval is much larger than the filters applied to each frame.

The horizontal blanking interval must also meet these minimums:

• If the kernel size is less than 4, and you use edge padding, the total porch must be at least 8 pixels.

• When you disable edge padding, the horizontal blanking interval must be at least 12 cycles and is independent of the kernel size.

• The BackPorch must be at least 6 pixels. This parameter is the number of inactive pixels before the first valid pixel in a frame.

For more information, see Configure Blanking Intervals.

Performance

This table shows the resource use after synthesis of the block for the Xilinx^® Zynq^®-7000 SoC ZC706 Evaluation Kit with single-pixel uint8 input and the default parameter settings. The design achieves a clock frequency of 325 MHz.

Resource	Usage
Slice LUTs	332
Slice Registers	627
DSP48	1
Block RAM	1

Extended Capabilities

C/C++ Code Generation

Generate C and C++ code using Simulink® Coder™.

This block supports C/C++ code generation for Simulink accelerator and rapid accelerator modes and for DPI component generation.

HDL Code Generation

Generate VHDL, Verilog and SystemVerilog code for FPGA and ASIC designs using HDL Coder™.

HDL Coder™ provides additional configuration options that affect HDL implementation and synthesized logic.

HDL Architecture

This block has one default HDL architecture.

HDL Block Properties

ConstrainedOutputPipeline	Number of registers to place at the outputs by moving existing delays within your design. Distributed pipelining does not redistribute these registers. The default is 0. For more details, see ConstrainedOutputPipeline (HDL Coder).
InputPipeline	Number of input pipeline stages to insert in the generated code. Distributed pipelining and constrained output pipelining can move these registers. The default is 0. For more details, see InputPipeline (HDL Coder).
OutputPipeline	Number of output pipeline stages to insert in the generated code. Distributed pipelining and constrained output pipelining can move these registers. The default is 0. For more details, see OutputPipeline (HDL Coder).

Version History

Introduced in R2017b

R2022a: Multipixel streaming

The Bilateral Filter block now supports multipixel streams. The HDL implementation replicates the algorithm for each pixel in parallel.

The block supports input column vectors of NumPixels values, where NumPixels can be 2, 4, or 8. The ctrl ports remain scalar, and the control signals in the pixelcontrol bus apply to all pixels in the matrix. The block returns an output vector of the same size as the input vector.

R2021b: Reflection padding

Pad the edge of a frame by reflecting around the edge-pixel value. This padding method helps reduce edge contrast effects and can improve results for machine learning while maintaining the original frame size.

R2020a: Option to omit padding

You can now configure the block to not add padding around the boundaries of the active frame. This option reduces the hardware resources used by the block and the blanking required between frames but affects the accuracy of the output pixels at the edges of the frame. To use this option, set the Padding method parameter to None. For an example, see Increase Throughput by Omitting Padding.

R2018b: Improved line buffer

The internal line buffer in this block now handles bursty data, that is, noncontiguous valid signals within a pixel line. This implementation uses fewer hardware resources due to improved padding logic and native support for kernel sizes with an even number of lines. This change affects the Line Buffer block and blocks that use an internal line buffer.

The latency of the line buffer is now reduced by a few cycles for some configurations. You might need to rebalance parallel path delays in your models. A best practice is to synchronize parallel paths in your models by using the pixel stream control signals rather than by inserting a specific number of delays.