Enhanced Object Segmentation for Vehicle Tracking and Dental CBCT by Neuromorphic Visual Processing with Controlled Neuron

Abstract

The neuromorphic visual processing is inspired by the robust visual recognition of human brain for the robust computer vision in ordinary everyday environment, by mimicking the behavior of primary visual cortex. With the recent wide application of deep neural networks approach for pattern recognition and artificial intelligence, the proposed neuromorphic neural network of visual processing was analyzed for the ADAS and the enhanced road safety. The feasibility of proposed neuromorphic design methodology with controlled neurons is demonstrated for the object segmentation technology for the vehicle tracking at night time and the tooth segmentation of noisy dental x-ray image via maintaining the successful segmentation without complicated post-processing layers or supervised learning. The new neuron integrated with both the rectifier and the resizing enhanced the performance via the successful segmentation over 97% with the night-time road traffic, without the post-processing denoising network layer. The post enhancement or integration with further deep networks becomes more flexible from incorporating the new neuron of rectifier and resizing demonstrating the abstracting efficiency, and the simple structure has the advantages of real-time and robust neuromorphic vision implemented by the small embedded system.