Abstract

Deep networks have already shown a spectacular success for object classification and detection for various applications from everyday use cases to advanced medical problems. The main advantage of the classification models over the detection models is less time and effort needed for dataset preparation, because classification networks do not require bounding box annotations, but labels at the image level only. Yet, after passing the image through a stack of convolutions followed by stride and pooling operations the full image is reduced into a single vector of class probabilities and the spatial arrangement of pixels is completely lost. Our proposed approach shows how to localize objects by restoring the spatial information about features distribution from the classification network. The presented modifications of architecture are limited to the inference phase only. As a result, we are able to combine the simplicity and short time required for dataset preparation with the results similar to the output from the detection model. Additionally, we showcased that it is possible to repurpose the existing network, originally trained on RGB images, to a novel task of facial features detection from low resolution thermal images, while preserving the high precision (>99%). The proposed facial features detection system can be potentially combined with wearable devices that will collect data and send it to the server for more computationally expensive analysis, e.g., calculation of the respiratory rate form the detected nostril area. Real time performance and small resource utilization proved that the presented approach can be used to serve multiple patients without any impact on the latency, e.g., as a centralized monitoring station for the home healthcare. The combination of wearable devices with machine learning algorithms run on the remote and more powerful platforms could revolutionize the practice of medicine by delivering healthcare to patients anywhere in the world.

Citations

Authors (4)