Deep Learning to Enhance Investigative Lead Information for Automotive Clear Coats

Currently, modern automotive paints consist of a thin undercoat and color coat layer protected by a thicker clear coat layer. All too often, a clear coat paint smear is the only layer of automotive paint left at the scene of a \"hit-and-run\" where damage to a vehicle and/or injury or death to a pedestrian has occurred. In these situations, searches of an automotive paint database such as PDQ database would generate thousands of hits as the coding in PDQ is generic and would lead to non-specific search criteria resulting in an extraordinarily large number of samples that a scientist must then work through and eliminate. Clear coat formulations are too similar for commercial library search algorithms to generate accurate hit lists by searching clear coat IR spectra alone. The inability of FTIR spectroscopy and the PDQ database to identify a vehicle from a clear coat paint smear using the text based searching system of PDQ or standard spectral library search algorithms is a significant limitation in the use of FTIR spectroscopy and forensic automotive paint databases to obtain information about the manufacturer, make, model, and production year of the vehicle from an automotive paint sample. To assess the evidential information content of the clear coat layer and improve upon our results of previously published studies on infrared library searching of automotive clear coats, we propose to develop a deep learning system using a convolutional neural network (CNN) to create a non-linear embedding of the data in a shared semantic feature space. The shared feature space can then be used by the neural networks for manufacturer classification, assembly plant classification, and for sample similarity matching. The CNN architecture will allow for a simultaneous optimization of and a transfer of learning between the different end components of the system. The shared feature space will make it easy to extend the architecture and is readily adaptable to new data and new domain knowledge. Although CNN training will be time consuming, the application of the resulting neural network is seamless and requires little in the way of computational resources. The output of deep learning classification will provide a conditional probability estimate. In addition, the shared embedded feature space can be used to create probability estimates using either nearest neighbor or Parzen window methods. Combining these two estimate will increase the confidence in and provide an understanding of the match.