From Open-Source ONNX Models to Aircraft Detection: What’s New in ENVI Deep Learning 4.0
With the release of ENVI® Deep Learning 4.0, the process of detecting objects in high-resolution satellite imagery just became much easier. Here are a few highlights of this release:
-
It supports the Open Neural Network Exchange (ONNX) model format. ONNX is an open-source standard for training deep learning models.
-
It provides new network architectures and improved training workflows. An example is the Real-Time Detectron (RT-DETR V2) object detection architecture, which is faster and simpler than its predecessors.
-
You can connect to the Analytics Repository for seamless sharing, discovery, and reuse of deep-learning models.
The video connected to this blog demonstrates how I used an open-source training dataset of military aircraft to detect Russian aircraft in a SkySat true-color image (courtesy of Planet Labs). The RoboFlow Universe web site is an open-source repository containing thousands of trained models and training rasters from community contributors.
For this scenario, I acquired roughly 5,000 training images of U.S. and Russian aircraft from the Aircraft Computer Vision Dataset at RoboFlow Universe. I downloaded the collection in Common Objects in Context (COCO) format and used the new Build Object Detection Rasters from COCO tool in ENVI Deep Learning 4.0 to create label rasters.
I used the Train Deep Learning Object Model tool to create a trained model in ONNX format, using 80% of the label rasters for training and 20% for validation. Finally, I used the trained ONNX object detection model to detect Russian bombers in a SkySat image with 0.5-meter spatial resolution. This is one of the benefits of ENVI Deep Learning; you can train a model once and reuse it to classify other images that are spatially and spectrally similar. I used the full version of ENVI Deep Learning, which enables you to train your own deep learning models and process data with ONNX models.
To summarize, the video demonstrates how you can leverage open-source training datasets to build your own object detection model in ENVI Deep Learning. Hopefully it provides some inspiration for your next project!