Real-Time Location-to-Image Generative Adversarial Networks with Sparse Sensor Data

The prevailing machine vision-based system monitoring faces challenges in complex optical environments, such as insufficient lumination or over-exposed light, which results in a severe reduction in the accuracy of visual surveillance. To fill the gap, this work proposes a novel dual generative adversarial networks model for real-time location-to-image transformation from the sparse distributed positional sensors. The conditional generative adversarial network (cGANis applied to learn high-dimensional features from pre-generated images for enhanced accuracy. The cycleGAN, with improved structure loss and key point loss functions, further improves the generation accuracy. The effectiveness of the proposed method is verified by real monitoring experiments designed to collect a few training data for both sensor data and real images. Experimental results show that the proposed method outperforms the mainstream cGANs variants in terms of overall generation quality and detail edge sharpening.