CLI for Neural Network image classification using Coco model.
This package is written to be used as a Node command-line interface (CLI).
This package uses opencv4nodejs which downloads and compiles opencv into Node.
Make sure you have cmake installed and available on your path or the build will fail.
To verify cmake, run:
command -v cmake
If available, you should see something like this:
$ command -v cmake /usr/bin/cmake
If it is not available, then it will be like this:
$ command -v cmake $
Apparently, this project is too big for
out of heap error when trying to upload it to the registry.
So, instead of:
npm i -g classify
You need to download the project from Github (this may be lengthy because of the size). Then run:
npm i -g . classify
To install the package globally.
This installs it globally into your Node ecosystem and makes it available on your path.
Now you can run it like so:
classify --image <path to image> --confidence 50
This is my first foray into Classification with Neural Networks. Another programmer at work did something similar in Python. I wanted to know if it was at all possible to do the same thing with Node. I found tensorflow.js and then tfjs-node (tensorflow for node), but had issues getting models converted to a web-friendly format for it to work. Then I found opencv4nodejs and this article Node.js meets OpenCV's Deep Neural Networks -- Fun with Tensorflow and Caffe. After that things fell into place. This cli project is the results of that endeavor. Feel free to add PRs if you would like it updated.
I am by no means an expert in this area. I am still learning (and there is a LOT to learn!). If asked, it's unlikely I'll be able to answer specific questions about Neural Networks or OpenCV.
Issues (large files)
Aside from the programming of the CLI, the biggest issue I found was uploading the models to Github. In order to have large file support, you must install git-lfs. I still ran into problems pushing the large files (being asked for username/password, which would fail), until I found that you also have to do this:
git remote set-url origin firstname.lastname@example.org:username/repo.git
It still took a (very!) long time to upload the project.
Running classify from the command-line will output the following:
classify Classifies an image using machine learning from passed in image path.
--image imagePath [required] The image path.
--confidence value [optional; default 50] The minimum confidence level to use for classification. (ex: 50 for 50%).
--filter filterFile [optional] A filter file used to filter out classification not wanted.
--quick [optional; default slow] Use quick classification, but may be more inaccurate.
--version Application version.
--help Print this usage guide.
--image or -i followed by the path to the image to classify.
--confidence or -c followed by the confidence value as a percentage (whole number). For instance, to filter out any levels less than 50%, use --confidence 50.
--filter or -f followed by a path to the filter file.
A filter file contains only the interested items from the model that you want classified. It contains one item per line. There should be no empty lines or comments.
bear bicycle bird bus car cat cow dog horse motorcycle person sheep train truck
--quick or -q specifies to use the 300x300 Coco SSD instead of the 512x512 Coco SSD. The 300x300 is faster, but may come at the cost of less accuracy in classified items.
--version or -v outputs the curent version.
--help or -h displays the help output.
The classified image will be output in the current directory. It is renamed in the following format:
bicycle_classified_coco300_30.jpg, with the latter containing rects and classification (if any exist).
classify --image ./images/test/bicycle.jpg --quick
classify --image ./images/test/train.jpg --confidence 50
classify --image ./images/test/royals.jpg --filter ./filter.txt
classify --image ./images/test/snapshot_001.jpg --filter ./filter.txt --confidence 50
Too Many Classifications
If you have an image with a lot of "action", consider filtering using the --filter parameter or at least the --confidence parameter.
This is what a classification looks like without either:
And, what it looks like after filtering: