研究人的视网膜并用于图像处理 [OpenCV]

目标

这篇文章主要呈现了一个人类视网膜模型，用于展示一些有趣图像处理和增强的特性。在这篇文章中你将学到：

• 从你的视网膜中发掘两个主通道
• 视网膜模型的基本使用
• 视网膜处理的一些参数调整

总体概述

该模型源于 Jeanny Herault 在 Gipsa 的研究，这是一个关于使用 Listic (code
maintainer) 进行图像处理的实验室。这并非完整的模型，但是已经可以呈现一些有趣的事情，这些可以用于增强图像的处理体验。该模型可以使用人类的视网膜信息：

• spectral whitening that has 3 important effects: high spatio-temporal frequency signals canceling (noise), mid-frequencies details enhancement and low frequencies luminance energy reduction. This all in one property directly
  allows visual signals cleaning of classical undesired distortions introduced by image sensors and input luminance range.
• local logarithmic luminance compression allows details to be enhanced even in low light conditions.
• decorrelation of the details information (Parvocellular output channel) and transient information (events, motion made available at the Magnocellular output channel).

The first two points are illustrated below :

In the figure below, the OpenEXR image sample CrissyField.exr, a High Dynamic Range image is shown. In order to make it visible on this web-page, the original input image
is linearly rescaled to the classical image luminance range [0-255] and is converted to 8bit/channel format. Such strong conversion hides many details because of too strong local contrasts. Furthermore, noise energy is also strong and pollutes visual information.

In the following image, as your retina does, local luminance adaptation, spatial noise removal and spectral whitening work together and transmit accurate information on lower range 8bit data channels.
On this picture, noise in significantly removed, local details hidden by strong luminance contrasts are enhanced. Output image keeps its naturalness and visual content is enhanced.

Note : image sample can be downloaded from the OpenEXR
website. Regarding this demonstration, before retina processing, input image has been linearly rescaled within 0-255 keeping its channels float format. 5% of its histogram ends has been cut (mostly removes wrong HDR pixels). Check out the sample opencv/samples/cpp/OpenEXRimages_HDR_Retina_toneMapping.cpp for
similar processing. The following demonstration will only consider classical 8bit/channel images.

The retina model output channels

The retina model presents two outputs that benefit from the above cited behaviors.

• The first one is called the Parvocellular channel. It is mainly active in the foveal retina area (high resolution central vision with color sensitive photo-receptors), its aim is to provide accurate color vision for visual details remaining static
  on the retina. On the other hand objects moving on the retina projection are blurred.
• The second well known channel is the Magnocellular channel. It is mainly active in the retina peripheral vision and send signals related to change events (motion, transient events, etc.). These outing signals also help visual system to focus/center
  retina on ‘transient’/moving areas for more detailed analysis thus improving visual scene context and object classification.

NOTE : regarding the proposed model, contrary to the real retina, we apply these two channels on the entire input images using the same resolution. This allows
enhanced visual details and motion information to be extracted on all the considered images... but remember, that these two channels are complementary. For example, if Magnocellular channel gives strong energy in an area, then, the Parvocellular channel is
certainly blurred there since there is a transient event.

As an illustration, we apply in the following the retina model on a webcam video stream of a dark visual scene. In this visual scene, captured in an amphitheater of the university, some students
are moving while talking to the teacher.

In this video sequence, because of the dark ambiance, signal to noise ratio is low and color artifacts are present on visual features edges because of the low quality image capture tool-chain.

Below is shown the retina foveal vision applied on the entire image. In the used retina configuration, global luminance is preserved and local contrasts are enhanced. Also, signal to noise ratio
is improved : since high frequency spatio-temporal noise is reduced, enhanced details are not corrupted by any enhanced noise.

Below is the output of the Magnocellular output of the retina model. Its signals are strong where transient events occur. Here, a student is moving at the bottom of the image thus generating high
energy. The remaining of the image is static however, it is corrupted by a strong noise. Here, the retina filters out most of the noise thus generating low false motion area ‘alarms’. This channel can be used as a transient/moving areas detector : it would
provide relevant information for a low cost segmentation tool that would highlight areas in which an event is occurring.

Retina use case

This model can be used basically for spatio-temporal video effects but also in the aim of :

• performing texture analysis with enhanced signal to noise ratio and enhanced details robust against input images luminance ranges (check out the Parvocellular retina channel output)
• performing motion analysis also taking benefit of the previously cited properties.

For more information, refer to the following papers :

• Benoit A., Caplier A., Durette B., Herault, J., “Using Human Visual System Modeling For Bio-Inspired Low Level Image Processing”, Elsevier, Computer Vision and Image Understanding 114 (2010), pp. 758-773. DOI <http://dx.doi.org/10.1016/j.cviu.2010.01.011>
• Please have a look at the reference work of Jeanny Herault that you can read in his book :

Vision: Images, Signals and Neural Networks: Models of Neural Processing in Visual Perception (Progress in Neural Processing),By: Jeanny Herault, ISBN: 9814273686. WAPI (Tower ID): 113266891.

This retina filter code includes the research contributions of phd/research collegues from which code has been redrawn by the author :

• take a look at the retinacolor.hpp module to discover Brice Chaix de Lavarene phD color mosaicing/demosaicing and his reference paper: B. Chaix de Lavarene, D. Alleysson, B. Durette, J. Herault (2007). “Efficient demosaicing
  through recursive filtering”, IEEE International Conference on Image Processing ICIP 2007
• take a look at imagelogpolprojection.hpp to discover retina spatial log sampling which originates from Barthelemy Durette phd with Jeanny Herault. A Retina / V1 cortex projection is also proposed and originates from Jeanny’s
  discussions. ====> more information in the above cited Jeanny Heraults’s book.

Code tutorial

Please refer to the original tutorial source code in file opencv_folder/samples/cpp/tutorial_code/contrib/retina_tutorial.cpp.

To compile it, assuming OpenCV is correctly installed, use the following command. It requires the opencv_core (cv::Mat and friends objects management), opencv_highgui (display
and image/video read) and opencv_contrib (Retina description) libraries to compile.

// compile
gcc retina_tutorial.cpp -o Retina_tuto -lopencv_core -lopencv_highgui -lopencv_contrib

// Run commands : add 'log' as a last parameter to apply a spatial log sampling (simulates retina sampling)
// run on webcam
./Retina_tuto -video
// run on video file
./Retina_tuto -video myVideo.avi
// run on an image
./Retina_tuto -image myPicture.jpg
// run on an image with log sampling
./Retina_tuto -image myPicture.jpg log

下面是代码的解释：

视网膜定义出现在 contrib 包中，并包含一个简单的使用示例：

#include "opencv2/opencv.hpp"

通过一个 help 函数为用户提供一些运行的提示：

// the help procedure
static void help(std::string errorMessage)
{
 std::cout<<"Program init error : "<<errorMessage<<std::endl;
 std::cout<<"\nProgram call procedure : retinaDemo [processing mode] [Optional : media target] [Optional LAST parameter: \"log\" to activate retina log sampling]"<<std::endl;
 std::cout<<"\t[processing mode] :"<<std::endl;
 std::cout<<"\t -image : for still image processing"<<std::endl;
 std::cout<<"\t -video : for video stream processing"<<std::endl;
 std::cout<<"\t[Optional : media target] :"<<std::endl;
 std::cout<<"\t if processing an image or video file, then, specify the path and filename of the target to process"<<std::endl;
 std::cout<<"\t leave empty if processing video stream coming from a connected video device"<<std::endl;
 std::cout<<"\t[Optional : activate retina log sampling] : an optional last parameter can be specified for retina spatial log sampling"<<std::endl;
 std::cout<<"\t set \"log\" without quotes to activate this sampling, output frame size will be divided by 4"<<std::endl;
 std::cout<<"\nExamples:"<<std::endl;
 std::cout<<"\t-Image processing : ./retinaDemo -image lena.jpg"<<std::endl;
 std::cout<<"\t-Image processing with log sampling : ./retinaDemo -image lena.jpg log"<<std::endl;
 std::cout<<"\t-Video processing : ./retinaDemo -video myMovie.mp4"<<std::endl;
 std::cout<<"\t-Live video processing : ./retinaDemo -video"<<std::endl;
 std::cout<<"\nPlease start again with new parameters"<<std::endl;
 std::cout<<"****************************************************"<<std::endl;
 std::cout<<" NOTE : this program generates the default retina parameters file 'RetinaDefaultParameters.xml'"<<std::endl;
 std::cout<<" => you can use this to fine tune parameters and load them if you save to file 'RetinaSpecificParameters.xml'"<<std::endl;
}

Then, start the main program and first declare a cv::Mat matrix in which input images will be loaded. Also allocate a cv::VideoCapture object ready
to load video streams (if necessary)

int main(int argc, char* argv[]) {
  // declare the retina input buffer... that will be fed differently in regard of the input media
  cv::Mat inputFrame;
  cv::VideoCapture videoCapture; // in case a video media is used, its manager is declared here

In the main program, before processing, first check input command parameters. Here it loads a first input image coming from a single loaded image (if user chose command -image)
or from a video stream (if user chose command -video). Also, if the user added log command at the end of its program call, the spatial logarithmic image sampling performed by the retina is taken into account by the
Boolean flag useLogSampling.

// welcome message
  std::cout<<"****************************************************"<<std::endl;
  std::cout<<"* Retina demonstration : demonstrates the use of is a wrapper class of the Gipsa/Listic Labs retina model."<<std::endl;
  std::cout<<"* This demo will try to load the file 'RetinaSpecificParameters.xml' (if exists).\nTo create it, copy the autogenerated template 'RetinaDefaultParameters.xml'.\nThen twaek it with your own retina parameters."<<std::endl;
  // basic input arguments checking
  if (argc<2)
  {
      help("bad number of parameter");
      return -1;
  }

  bool useLogSampling = !strcmp(argv[argc-1], "log"); // check if user wants retina log sampling processing

  std::string inputMediaType=argv[1];

  //////////////////////////////////////////////////////////////////////////////
  // 检查输入的媒体类型 (静态图片、视频文件以及现场视频采集)
  if (!strcmp(inputMediaType.c_str(), "-image") && argc >= 3)
  {
      std::cout<<"RetinaDemo: processing image "<<argv[2]<<std::endl;
      // image processing case
      inputFrame = cv::imread(std::string(argv[2]), 1); // load image in BGR color mode
  }else
      if (!strcmp(inputMediaType.c_str(), "-video"))
      {
          if (argc == 2 || (argc == 3 && useLogSampling)) // attempt to grab images from a video capture device
          {
              videoCapture.open(0);
          }else// attempt to grab images from a video filestream
          {
              std::cout<<"RetinaDemo: processing video stream "<<argv[2]<<std::endl;
              videoCapture.open(argv[2]);
          }

          // grab a first frame to check if everything is ok
          videoCapture>>inputFrame;
      }else
      {
          // bad command parameter
          help("bad command parameter");
          return -1;
      }

Once all input parameters are processed, a first image should have been loaded, if not, display error and stop program :

if (inputFrame.empty())
{
    help("Input media could not be loaded, aborting");
    return -1;
}

Now, everything is ready to run the retina model. I propose here to allocate a retina instance and to manage the eventual log sampling option. The Retina constructor expects at least a cv::Size
object that shows the input data size that will have to be managed. One can activate other options such as color and its related color multiplexing strategy (here Bayer multiplexing is chosen using enum cv::RETINA_COLOR_BAYER). If using log sampling, the image
reduction factor (smaller output images) and log sampling strengh can be adjusted.

// pointer to a retina object
cv::Ptr<cv::Retina> myRetina;

// if the last parameter is 'log', then activate log sampling (favour foveal vision and subsamples peripheral vision)
if (useLogSampling)
{
    myRetina = new cv::Retina(inputFrame.size(), true, cv::RETINA_COLOR_BAYER, true, 2.0, 10.0);
}
else// -> else allocate "classical" retina :
    myRetina = new cv::Retina(inputFrame.size());

一旦完成，上述代码就会生成一个默认 xml 文件，包含视网膜的默认参数。这个可以让你很方便的在你的配置中使用这个模板。这里所生成的模板 xml 文件名为 RetinaDefaultParameters.xml.

// 保存默认视网膜参数文件，以便阅读并修改然后重载
myRetina->write("RetinaDefaultParameters.xml");

紧接着，程序视图加载另外一个名为 RetinaSpecificParameters.xml  的 XML 文件文件。如果你已经创建了这个文件并引入，那么该文件会被加载，否则将使用默认的视网膜参数。

// 如果文件存在则加载
myRetina->setup("RetinaSpecificParameters.xml");

这个文件并不是必须的，只是提供了这样一个选择，你可以清空视网膜缓冲区来强制清除以往事件。

// 重置所有视网膜缓冲区 (想想你闭上眼睛很长时间)
myRetina->clearBuffers();

现在可以运行视网膜程序了，首先我创建了一个输出缓冲区用于接收两个视网膜频道的输出：

// declare retina output buffers
cv::Mat retinaOutput_parvo;
cv::Mat retinaOutput_magno;

然后在循环中运行视网膜，根据需要从视频序列中加载新的帧，并将输出写回专用缓冲区。

// 无限循环
while(true)
{
    // if using video stream, then, grabbing a new frame, else, input remains the same
    if (videoCapture.isOpened())
        videoCapture>>inputFrame;

    // run retina filter on the loaded input frame
    myRetina->run(inputFrame);
    // Retrieve and display retina output
    myRetina->getParvo(retinaOutput_parvo);
    myRetina->getMagno(retinaOutput_magno);
    cv::imshow("retina input", inputFrame);
    cv::imshow("Retina Parvo", retinaOutput_parvo);
    cv::imshow("Retina Magno", retinaOutput_magno);
    cv::waitKey(10);
}

这样就完成了。但是如果你想确保系统安全，请小心管理异常。该程序在一些无效数据会发生异常，例如没有输入帧、错误的设置等等。接下来我建议使用 try catch 代码块来执行视网膜处理代码，如下所示：

try{
     // pointer to a retina object
     cv::Ptr<cv::Retina> myRetina;
     [---]
     // processing loop with no stop condition
     while(true)
     {
         [---]
     }

}catch(cv::Exception e)
{
    std::cerr<<"Error using Retina : "<<e.what()<<std::endl;
}

视网膜参数该如何处理？

首先，推荐如下参考论文：

• Benoit A., Caplier A., Durette B., Herault, J., “Using Human Visual System Modeling For Bio-Inspired Low Level Image Processing”, Elsevier, Computer Vision and Image Understanding 114 (2010), pp. 758-773. DOI <http://dx.doi.org/10.1016/j.cviu.2010.01.011>

然后打开上述示例代码生成的配置文件 RetinaDefaultParameters.xml ，内容如下：

<?xml version="1.0"?>
<opencv_storage>
<OPLandIPLparvo>
  <colorMode>1</colorMode>
  <normaliseOutput>1</normaliseOutput>
  <photoreceptorsLocalAdaptationSensitivity>7.0e-01</photoreceptorsLocalAdaptationSensitivity>
  <photoreceptorsTemporalConstant>5.0e-01</photoreceptorsTemporalConstant>
  <photoreceptorsSpatialConstant>5.3e-01</photoreceptorsSpatialConstant>
  <horizontalCellsGain>0.</horizontalCellsGain>
  <hcellsTemporalConstant>1.</hcellsTemporalConstant>
  <hcellsSpatialConstant>7.</hcellsSpatialConstant>
  <ganglionCellsSensitivity>7.0e-01</ganglionCellsSensitivity></OPLandIPLparvo>
<IPLmagno>
  <normaliseOutput>1</normaliseOutput>
  <parasolCells_beta>0.</parasolCells_beta>
  <parasolCells_tau>0.</parasolCells_tau>
  <parasolCells_k>7.</parasolCells_k>
  <amacrinCellsTemporalCutFrequency>1.2e+00</amacrinCellsTemporalCutFrequency>
  <V0CompressionParameter>9.5e-01</V0CompressionParameter>
  <localAdaptintegration_tau>0.</localAdaptintegration_tau>
  <localAdaptintegration_k>7.</localAdaptintegration_k></IPLmagno>
</opencv_storage>

Here are some hints but actually, the best parameter setup depends more on what you want to do with the retina rather than the images input that you give to retina. Apart from the more specific
case of High Dynamic Range images (HDR) that require more specific setup for specific luminance compression objective, the retina behaviors should be rather stable from content to content. Note that OpenCV is able to manage such HDR format thanks to the OpenEXR
images compatibility.

Then, if the application target requires details enhancement prior to specific image processing, you need to know if mean luminance information is required or not. If not, the the retina can cancel
or significantly reduce its energy thus giving more visibility to higher spatial frequency details.

Basic parameters

The most simple parameters are the following :

• colorMode : let the retina process color information (if 1) or gray scale images (if 0). In this last case, only the first channel of the input will be processed.
• normaliseOutput : each channel has this parameter, if value is 1, then the considered channel output is rescaled between 0 and 255. Take care in this case at the Magnocellular output level (motion/transient
  channel detection). Residual noise will also be rescaled !

Note : using color requires color channels multiplexing/demultipexing which requires more processing. You can expect much faster processing using gray levels
: it would require around 30 product per pixel for all the retina processes and it has recently been parallelized for multicore architectures.

Photo-receptors parameters

The following parameters act on the entry point of the retina - photo-receptors - and impact all the following processes. These sensors are low pass spatio-temporal filters that smooth temporal
and spatial data and also adjust there sensitivity to local luminance thus improving details extraction and high frequency noise canceling.

• photoreceptorsLocalAdaptationSensitivity between 0 and 1. Values close to 1 allow high luminance log compression effect at the photo-receptors level. Values closer to 0 give a more linear sensitivity. Increased
  alone, it can burn the Parvo (details channel) output image. If adjusted in collaboration with ganglionCellsSensitivity images can be very contrasted whatever the local luminance there is... at the
  price of a naturalness decrease.
• photoreceptorsTemporalConstant this setups the temporal constant of the low pass filter effect at the entry of the retina. High value lead to strong temporal smoothing effect : moving objects are blurred and
  can disappear while static object are favored. But when starting the retina processing, stable state is reached lately.
• photoreceptorsSpatialConstant specifies the spatial constant related to photo-receptors low pass filter effect. This parameters specify the minimum allowed spatial signal period allowed in the following. Typically,
  this filter should cut high frequency noise. Then a 0 value doesn’t cut anything noise while higher values start to cut high spatial frequencies and more and more lower frequencies... Then, do not go to high if you wanna see some details of the input images
  ! A good compromise for color images is 0.53 since this won’t affect too much the color spectrum. Higher values would lead to gray and blurred output images.

Horizontal cells parameters

This parameter set tunes the neural network connected to the photo-receptors, the horizontal cells. It modulates photo-receptors sensitivity and completes the processing for final spectral whitening
(part of the spatial band pass effect thus favoring visual details enhancement).

• horizontalCellsGain here is a critical parameter ! If you are not interested by the mean luminance and focus on details enhancement, then, set to zero. But if you want to keep some environment luminance data,
  let some low spatial frequencies pass into the system and set a higher value (<1).
• hcellsTemporalConstant similar to photo-receptors, this acts on the temporal constant of a low pass temporal filter that smooths input data. Here, a high value generates a high retina after effect while a lower
  value makes the retina more reactive.
• hcellsSpatialConstant is the spatial constant of the low pass filter of these cells filter. It specifies the lowest spatial frequency allowed in the following. Visually, a high value leads to very low spatial
  frequencies processing and leads to salient halo effects. Lower values reduce this effect but the limit is : do not go lower than the value ofphotoreceptorsSpatialConstant. Those 2 parameters actually specify the spatial
  band-pass of the retina.

NOTE after the processing managed by the previous parameters, input data is cleaned from noise and luminance in already partly enhanced. The following parameters
act on the last processing stages of the two outing retina signals.

Parvo (details channel) dedicated parameter

• ganglionCellsSensitivity specifies the strength of the final local adaptation occurring at the output of this details dedicated channel. Parameter values remain between 0 and 1. Low value tend to give a linear
  response while higher values enforces the remaining low contrasted areas.

Note : this parameter can correct eventual burned images by favoring low energetic details of the visual scene, even in bright areas.

IPL Magno (motion/transient channel) parameters

Once image information is cleaned, this channel acts as a high pass temporal filter that only selects signals related to transient signals (events, motion, etc.). A low pass spatial filter smooths
extracted transient data and a final logarithmic compression enhances low transient events thus enhancing event sensitivity.

• parasolCells_beta generally set to zero, can be considered as an amplifier gain at the entry point of this processing stage. Generally set to 0.
• parasolCells_tau the temporal smoothing effect that can be added
• parasolCells_k the spatial constant of the spatial filtering effect, set it at a high value to favor low spatial frequency signals that are lower subject to residual noise.
• amacrinCellsTemporalCutFrequency specifies the temporal constant of the high pass filter. High values let slow transient events to be selected.
• V0CompressionParameter specifies the strength of the log compression. Similar behaviors to previous description but here it enforces sensitivity of transient events.
• localAdaptintegration_tau generally set to 0, no real use here actually
• localAdaptintegration_k specifies the size of the area on which local adaptation is performed. Low values lead to short range local adaptation (higher sensitivity to noise), high values secure log compression.