Programming Comments - Artificial Neural Networks

Feedback I received after training neural networks for clients to embed into their commercial C++ applications:

Client's Feedback To You
Five Stars!

Pure Genius. Stephane and his tools DarkHelp and
DarkMark are the best thing that happened to us
and helped us take a leap into deep learning .. He
knows the details and he documents things in a
crisp and clear manner. Very wonderful to work
with, always available and informative ... He went
beyond the scope of the project to make sure we
got everything working on the hardware we
wanted... Highly recommended!

Very knowledgeable and responsive AI/ML engineer, very
professional, a pleasure to work with. Would love to work
with him again!

The tools I use to train and embed artificial neural networks into C++ applications are Darknet[1], DarkHelp[2], and DarkMark.

All of these neural networks run in C++ using the DarkHelp C++ neural network wrapper library.
Image markup to train these networks was done using DarkMark, an open-source tool I designed and wrote specifically to help train artificial neural networks.
The libdarknet.so C library is 1.1 MiB in size, and the C++ wrapper libdarkhelp.so is 109 KiB in size.
- At a cost of just over 1 MiB, any C++ application can easily make use of an artificial neural network trained with Darknet.
- This makes it extremely trivial to use in a desktop application, and relatively simple to use even within an embedded device with limited resources, such as a low-cost Raspberry Pi or NVIDIA Jetson Nano.
For details on how this is done, see the stop sign tutorial I wrote.

Project	Notes	Samples Images
DarkPlate	DarkPlate is a C++ project for reading license plates. It uses Darknet/YOLO, DarkHelp, and OpenCV to find and parse vehicle license plates.
MoveDetect	MoveDetect is a C++ library to detect movement between frames in a video. It is very low impact, designed to run on tiny devices such as Raspberry Pi, Beaglebone, and Jetson Nano.
Pharmacy	A neural network trained to detect and count different kinds of pills. Works even when pills are on their side or stacked, and can highlight broken pills.
Fire	What if we could train our home and business security cameras to detect certain things like fire and contact us when those events are detected?
Sudoku	Using a YOLOv4-Tiny neural network to take apart a Sudoku, and then extracting the parts from the image using OpenCV to find the solution.
Counting Objects	Counting wooden dowels on a conveyer belt.
Stop Signs	A detailed stop sign tutorial I wrote on how to get started with Darknet, DarkHelp, and DarkMark. To help people get started with Darknet and DarkMark, I also recorded a video showing how to do the stop sign tutorial.	Click on the image above to view the gallery, or click below to view on Youtube.
Driving	Recognizing vehicles, bicycles, and pedestrians. DarkHelp has new support for annotating videos, not just JPEG and PNG files. This video was done as a demonstration of applying a neural network to video files. All from within C++ using a dozen lines of code and a library that is ~300 KiB in size! The option to display the processing time of each frame was enabled. In this video, you can see in the upper-left corner that each 1280x720 frame took 2 milliseconds to process. This gives a possible frame rate of ~500 FPS when processing videos offline.
Cancer Cells	Custom artificial neural network trained to recognize 15 different types of cancer cells: Basophil Eosinophil Erythroblast Lymphocyte (atypical) Lymphocyte (typical) Metamyelocyte Monoblast Monocyte Myeloblast Myelocyte Neutrophil (band) Neutrophil (segmented) Promyelocyte Promyelocyte (bilobled) Smudge cell Image source: A Single-cell Morphological Dataset of Leukocytes from TCIA (The Cancer Image Archive).	Click on the image to view the gallery.
YOLOv3 and YOLOv4	Comparison between the following 4 neural networks configurations: yolov3-tiny.cfg (2 layer) yolov3-tiny_3l.cfg (3 layer) yolov4-tiny.cfg (2 layer) yolov4-tiny-3l.cfg (3 layer) This video compares images from 4 similar neural networks, all trained the exact same way.
Mailboxes	Custom artificial neural network trained to recognize 17 classes related to Canada Post mail boxes: numbers 1 to 16 lock Image source: self.	Click on the image to view the gallery.
Handwashing	Handwashing recognition using a neural network. Each frame takes a decent GPU between 1 and 2 milliseconds to analyze, allowing video to be processed at > 300 FPS when taking into account Darknet and OpenCV overhead. Neural network and C++ library is built using Darknet, YOLOv3-tiny, DarkMark, and DarkHelp C++. See the open-source C++ handwashing library I published for details.
Street Signs	Custom artificial neural network trained to recognize 5 classes related to common street signs: stop sign yield sign street name speed limit back of stop sign Image source: self.	Click on the image to view the gallery.
Prairie Dogs	I had to train a neural network to find prairie dogs for some tests I was doing with OpenCV's CSRT object tracking. The tests were not successful, but the neural network remains.
Store Price Labels	Custom artificial neural network trained to recognize 12 classes related to price tags: digits 0 to 9 barcode description Image source: self.	Click on the image to view the gallery.
CSRT Tracking	I looked into adding CSRT object tracking into DarkMark. Interesting results, but I'm not convinced there would be any benefit to add full support for CSRT to DarkMark.
Dog Versus Cat	Custom artificial neural network trained to recognize just 2 classes: dogs cats See also: Dogs And Cats, a post I wrote with some interesting results during the training of this neural network. Image source: Kaggle "Dogs and Cats" project.	Click on the image to view the gallery.
Video	Custom artificial neural network trained to recognize 7 different classes, which is then applied to a video stream: eye, nose, mouth, ear, bottle, blue pen, orange pen YOLOv3 neural network was trained using the Darknet framework. Training (10K iterations) took 1.5 hours. Weight files are 33 MiB in size. Video capture and neural network image processing is written in C++. In this video, it is running in a VirtualBox VM without a GPU at ~7 FPS.
Iranian Plates	Automatic number-plate recognition (aka ANPR, ALPR, VLPR) for Iranian license plates.	Click on the image to view the gallery.