FeatureX is disrupting the field of satellite image analysis, making it easy for companies like yours to gain insights from satellite imagery. Our semantic deep learning technology enables our systems to understand the fundamental structure of images and then reimagine them in new ways, so you can extract more information from every image.
We do the heavy lifting: satellite image sourcing and licensing, image enhancement and alignment, and computer vision-based image understanding. Our technology enables you to focus on your business needs.
Images copyright 2017 DigitalGlobe, Inc.
Want to count the percentage of pine trees in a forest? How about cars in a rental car parking lot?
FeatureX's semantic object recognition algorithms provide you with unmatched tools to recognize small objects in satellite images. We can accurately identify objects as small as 5 x 5 pixels in high-resolution satellite imagery.
Building and Road Detection
We can detect buildings, find their shapes, and estimate the square footage of their footprint. We can also detect roads, both paved and non-paved. For parking lots, we can detect vehicle access to parking places.
In these examples, buildings are outlined in red, and roads and parking lots highlighted in blue.
Since we operate at the semantic level, we can take a grayscale image, and guided by low resolution color information, re-imagine the entire image at a higher resolution.
In the example here, a single original color pixel is replaced by 64 new color pixels. The re-imagined images are sharp and accurate, with the colors naturally balanced.
A common challenge in satellite imagery is pan sharpening - combining high resolution grey scale images with lower resolution color images to produce high resolution color images.
Our semantic approach allows our systems to understand both natural and artificial structure in images, producing stunning results that far surpass the competition.
Infrared Image Enhancement
Our ability to enhance images doesn’t stop with the visible spectrum. For example, near infrared can highlight details in images to facilitate tasks such as detection and measurement of ground cover and foliage. In these examples, each original infrared pixel is replaced with 16 high-resolution pixels.
Contrary to most computer vision problems, with satellite imagery there isn't really an up or down. Image orientation is at any angle, as are the angles of the sun, shadows, and satellite viewing angles. With conventional convolutional neural networks (CNNs), the orientation of an image for recognition must be roughly the same angle as what the system was trained with for recognition to be effective.
Our technology, FXnet, is insensitive to orientation. Once FXnet has been trained for one orientation, it can recognize objects at all other orientations without change.
In the example above, a standard high-performance CNN and FXnet were trained to recognize handwritten digits in the upright orientation. When presented with a rotating three, the CNN fails to recognize the three in most orientations. FXnet, on the other hand, recognizes the three at all orientations. To observe this, notice how stable the final fully connected layer is, as well as the softmax layer.
Being insensitive to image orientation provides a clear advantage. An added benefit is that FXnet can perform its magic with fewer parameters than an equivalent CNN, decreasing memory requirements.
Please contact us to discuss pricing for the services you need. We provide deep discounts for researchers and academic institutions, and we support non-profit organizations that monitor the health of our planet's environment.