Correlation using OpenCV

 This tutorial covers the correlation operator in OpenCV. The slides relevant to this tutorial can be downloaded from here. The source code for this tutorial can be accessed from here. The result files that we will discuss can be downloaded from here. 

Unzip the result and source zip files to a folder on your system and run the following commands: "cmake ..; make; ./homework". We have unzipped the source code into the folder /vision/practice/. You should see the following output:

acv@acv-vm:~/vision/practice/bin$ cmake ..; make; ./homework
-- The C compiler identification is GNU 7.5.0
-- The CXX compiler identification is GNU 7.5.0
-- Check for working C compiler: /usr/bin/cc
-- Check for working C compiler: /usr/bin/cc -- works
-- Detecting C compiler ABI info
-- Detecting C compiler ABI info - done
-- Detecting C compile features
-- Detecting C compile features - done
-- Check for working CXX compiler: /usr/bin/c++
-- Check for working CXX compiler: /usr/bin/c++ -- works
-- Detecting CXX compiler ABI info
-- Detecting CXX compiler ABI info - done
-- Detecting CXX compile features
-- Detecting CXX compile features - done
-- Found PkgConfig: /usr/bin/pkg-config (found version "0.29.1") 
-- Found OpenCV: /usr/local (found version "4.4.0") 
-- Configuring done
-- Generating done
-- Build files have been written to: /home/acv/vision/practice/bin
Scanning dependencies of target homework
[ 16%] Building CXX object CMakeFiles/homework.dir/src/CorrelationDemo.cpp.o
[ 33%] Building CXX object CMakeFiles/homework.dir/src/ImageUtilities.cpp.o
[ 50%] Building CXX object CMakeFiles/homework.dir/src/Morphology.cpp.o
[ 66%] Building CXX object CMakeFiles/homework.dir/src/homework.cpp.o
[ 83%] Building CXX object CMakeFiles/homework.dir/src/util/LogManager.cpp.o
[100%] Linking CXX executable homework
[100%] Built target homework


These files are used to submit homework for the ACV course.

Note - run the code from the "bin" folder:

Usage is as follows:

./homework rlerow inputimage.jpeg output_folder/ 

./homework dilate inputimage.jpeg output_folder/ 

./homework erode inputimage.jpeg output_folder/ 

./homework correlationDemo
Run the following code: "./homework correlationDemo". You should see the following files in your /tmp folder: 

/tmp/inImage_modified.png         /tmp/result_correlation.png
/tmp/kernelTemplate_modified.png  /tmp/result_overlay.png
The file "/tmp/result_overlay.png" must be exactly similar to the results file in "sample_results/result_1/result_overlay.png". Notice the image "sample_results/result_1/inImage_modified.png". This is a modified version of the input image "../sample_inputs/convolution/pcb_3.png". This image appears darker because we are correcting the image for the mean value in line no 24 of "CorrelationDemo.h". Comment this line, save the file and run the following commands: "make; ./homework correlationDemo". You should see the following lines in your command prompt: 


  
acv@acv-vm:~/vision/practice/bin$ make; ./homework correlationDemo
Scanning dependencies of target homework
[ 16%] Building CXX object CMakeFiles/homework.dir/src/CorrelationDemo.cpp.o
[ 33%] Building CXX object CMakeFiles/homework.dir/src/homework.cpp.o
[ 50%] Linking CXX executable homework
[100%] Built target homework
acv@acv-vm:~/vision/practice/bin$
The result file "/tmp/result_overlay.png" must be exactly similar to the results file in "sample_results/result_2/result_overlay.png". The component has not been detected correctly. The two result files are compared below. The first image has the correct result, while the image below it is the wrong result.
The correlation operator moves over the input image to different locations row i=0 to Height, column, j=0 to Width. At each location, the result is computed as a weighted sum of the input image with the template image according to the correlation forumla shown below.
Due to the weighted sum, uniformly bright areas in the input image get a high value irrespective of the template image used. Ensuring that the input and template image are normalized about the mean gives the correct result. You can experiment with other images given in the input set. Change line no 5 in CorrelationDemo.cpp to the following. 

  char inpImagePath[1024] = "../sample_inputs/convolution/pcb_3_blanked_out.png";
Save the file and run the following command: 

  
acv@acv-vm:~/vision/practice/bin$ make; ./homework correlationDemo
Scanning dependencies of target homework
[ 16%] Building CXX object CMakeFiles/homework.dir/src/CorrelationDemo.cpp.o
[ 33%] Linking CXX executable homework
[100%] Built target homework
acv@acv-vm:~/vision/practice/bin$
The result generated in /tmp/result_overlay.png" must be exactly similar to the results file in "sample_results/result_3/result_overlay.png". It is shown below:
Notice that the component has been matched with a white colored patch. Remove the comment from line no 24 of "CorrelationDemo.h". Save the file and run the following commands: "make; ./homework correlationDemo". You should see the following lines in your command prompt:  


  
acv@acv-vm:~/vision/practice/bin$ make; ./homework correlationDemo
Scanning dependencies of target homework
[ 16%] Building CXX object CMakeFiles/homework.dir/src/CorrelationDemo.cpp.o
[ 33%] Building CXX object CMakeFiles/homework.dir/src/homework.cpp.o
[ 50%] Linking CXX executable homework
[100%] Built target homework
acv@acv-vm:~/vision/practice/bin$
The result file generated in The file "/tmp/result_overlay.png" must be exactly similar to the results file in "sample_results/result_4/result_overlay.png". The result files are shown below for comparison. The first image is the mean-corrected input image. It is generated in line no 14 of CorrelationDemo.cpp. The lower image is the result from correlation matching :
Change the input image by modifying line no 5 of CorrelationDemo.cpp to the following: 

  char inpImagePath[1024] = "../sample_inputs/convolution/pcb_2.png";
Save the file and compile and run the code using following commands: 

acv@acv-vm:~/vision/practice/bin$ make; ./homework correlationDemo
Scanning dependencies of target homework
[ 16%] Building CXX object CMakeFiles/homework.dir/src/CorrelationDemo.cpp.o
[ 33%] Linking CXX executable homework
[100%] Built target homework
acv@acv-vm:~/vision/practice/bin$
The result file generated in The file "/tmp/result_overlay.png" must be exactly similar to the results file in "sample_results/result_5/result_overlay.png". It is reproduced below for reference. The component being matched is not present in the input image. Correlation has picked up the closest match.
However, this does not always happen. Sometimes correlation is unable to locate the nearest match for a template. For example, change the template image to "pcb_component_2.png" by modifying the line 6 in CorrelationDemo.cpp: 

  char templateImagePath[1024] = "../sample_inputs/convolution/pcb_component_2.png";
The result file generated in The file "/tmp/result_overlay.png" must be exactly similar to the results file in "sample_results/result_6/result_overlay.png". The component has NOT been detected even though very similar components are present on the input image. If you change the input image to "pcb_3.png", the component is detected. These results are in "sample_results/result_7/result_overlay.png". The plain vanilla correlation implemented here is not robust to scale and rotation variation.

Comments

Post a Comment

Popular posts from this blog

Discrete Wavelet Transform

Code and sample files for this demo can be downloaded from here . Code uses only libraries from the ACV-VirtualMachine.No extra libraries are needed. You can unzip the code to any folder on your system. All paths used in the code are relative. We are using the folder "~/acv/code/dwt_demo/" Compiling and running the code is fairly straight forward. You only need to run the following two commands: acv@acv-vm : ~/acv/code/dwt_demo/bin $ cmake ../src/wavelet_demo_others/ -- The C compiler identification is GNU 7.5.0 -- The CXX compiler identification is GNU 7.5.0 -- Check for working C compiler: /usr/bin/cc -- Check for working C compiler: /usr/bin/cc -- works -- Detecting C compiler ABI info -- Detecting C compiler ABI info - done -- Detecting C compile features -- Detecting C compile features - done -- Check for working CXX compiler: /usr/bin/c++ -- Check for working CXX compiler: /usr/bin/c++ -- works -- Detecting CXX compiler ABI info -- Detecting CXX compiler ABI info - ...
Read more

Geometric image transforms

Image
 The sample code and data for this tutorial can be downloaded from here . The result files are here . The accompanying slide deck for this tutorial is here . After downloading the code, run the following commands:  acv@acv-vm : ~/vision/practice/bin $ cmake ..; make; ./homework transformDemo -- The C compiler identification is GNU 7.5.0 -- The CXX compiler identification is GNU 7.5.0 -- Check for working C compiler: /usr/bin/cc -- Check for working C compiler: /usr/bin/cc -- works -- Detecting C compiler ABI info -- Detecting C compiler ABI info - done -- Detecting C compile features -- Detecting C compile features - done -- Check for working CXX compiler: /usr/bin/c++ -- Check for working CXX compiler: /usr/bin/c++ -- works -- Detecting CXX compiler ABI info -- Detecting CXX compiler ABI info - done -- Detecting CXX compile features -- Detecting CXX compile features - done -- Found PkgConfig: /usr/bin/pkg-config (found version "0.29.1") -- Found OpenCV: /usr/local (found...
Read more

Testing image capture with high speed cameras

Image
In industrial systems, objects may be moving along conveyor belts or carried on pick and place machines. A camera based inspection solution for such situations needs to take into accoun that the object in question may not be stationery, or at a fixed distance from the camera. Ergo, the captured image may not contain the object in focus, or not have the object in view at all. One option would be to stop the conveyor belt to take a photo of the object:   However, stopping a conveyor belt is not always possible. You would need cameras that can capture a clear image even when objects are moving. Enthusiasts naiively assume that one can take a video and perform analysis on individual frames of the video. Unless the video has the correct frame rate and good resolution, the solution would not work. Assume that you need to implement an industrial imaging solution to monitor items moving on a conveyor belt. You should consider specifications such as these: Distance from camera t...
Read more