FLIR Enhanced Vision

Project Type: 
Electrical & Computer Engineering
Year: 
2017

Files

Description

 

FLIR has recently released a miniature thermal camera called the Boson. The FLIR Boson camera utilizes a 320x256 thermal sensor as well as an onboard processor in a package that is significantly cheaper than an HD thermal camera. The downside with this camera is that the sensor is small, and finer details are missed upon inspection. This makes the Boson less applicable, less versatile, and less enticing to consumers. To combat this, an algorithm is being created that will upscale the image and attempt to improve its quality. The onboard processor is key here, as it will allow an algorithm to be implemented that will take a low resolution image from the sensor and make it appear as if it came from a much higher resolution sensor, but at no additional hardware cost. This processor is a 12 core CPU by Movidius called Myriad 2. This CPU is specially designed to be able to accelerate image processing and machine learning applications. To prevent this blurriness and quality degradation of the image when upscaling, a machine learning algorithm is being utilized to synthesize the additional information needed to create a higher quality image. This is an algorithm that will be trained to look for similarities between low resolution and high resolution images. This will allow one to synthesize high resolution pixel information from a low resolution input. To be a little more specific, a neural network will be used. This network will generate outputs by making decisions based on the characteristics of the input image, and performing certain calculations based on these decisions. The algorithm will enhance images by training a neural network with tens of thousands of thermal images. By the end of the training, the network will basically have the best probabilistic function that can be applied on any given picture to enlarge it. This function will try to predict and provide the additional information needed for an enlarged high definition output image.