The aim of Non-Destructive Evaluation (NDE) is to probe materials and structures to detect and characterize defects and discontinuities without disturbing the target material. In active thermography the material being investigated is subject to infrared radiation. The waveform of the stimulating radiation can take multiple forms including uniform radiation, pulsatile radiation of a sinusoidal input with a specific frequency. The thermal response of the material is monitored using an infrared camera. The spatiotemporal temperature field is analysed by suitable techniques to detect and classify the defects embedded in the material.

Although there are physics-based modelling approaches to the above problem, use of data-drive approaches based on deep neural networks has been gaining traction in the recent years. Particularly use of convolutional neural networks (CNNS) ideally suited for image pattern recognition has been explored. Although CNNs are ideal for classification of static images, they have been applied to classification of spatiotemporal data also by presenting a stack of frames as input to the CNN, thereby converting a temporal problem into a spatial problem. But such a conversion is unnatural. Ideally one must use a spatiotemporal classifier with appropriate learning algorithms.

We have developed a special variation of deep neural network architecture for temporal processing. The model consists of a layer of flip-flop neurons that are capable of retaining information for arbitrarily long times similar to the Long Short Term Memory (LSTM) networks. These flip-flop deep neural networks (FFDNNs) have been shown to yield excellent performance on the so-called long delay decision making problems. Use of specific kinds of flip-flops in the network enabled the network to solve specific kinds of temporal decision-making problems. For example, S-R flip flop neurons were found to be ideal for 1-2-AX-BY Task and use of Toggle flip-flop neurons, which was implemented using nonlinear oscillators, yielded excellent performance with the parity problem. Such models have also been able to explain a wide range of experimental phenomena in neuroscience.

The above line of work points to a novel line of temporal deep neural networks that can efficiently solve spatiotemporal data. In this project we propose to develop a hybrid of our FFDNNs and CNNs to create Flip-Flop CNNs (FFCNNs) and apply the same to the problem of localizing defects from active themographic images.