Modern automotive paint consists of a thin e-coat, primer and color coat layer protected by a thick clear coat layer. All too often, the clear coat is the only layer of automotive paint recovered at the crime scene of a “hit-and-run” where damage to a vehicle and/or injury or death to a pedestrian has occurred. In these cases, directly searching an automotive paint database such as the Royal Canadian Mounted Police paint data query (PDQ) database using commercial library search algorithm will generate a large number (thousands) of potential hits because of the large number of similar spectra. The inability of Fourier transform infrared (FTIR) spectroscopy and the PDQ database to identify the manufacturer and model of the vehicle from an automotive clear coat sample is a limitation in the use of FTIR spectroscopy and forensic automotive paint databases such as PDQ.
The authors previously demonstrated that pattern recognition, when applied directly to mid-infrared absorbance spectra of original equipment manufacturer (OEM) clear coats, has the potential to differentiate between similar clear coat IR spectra. The prototype pattern recognition-based infrared (IR) library search system consisted of two separate but interrelated components: search prefilters to reduce the size of the library of a specific manufacturer to an assembly plant or assembly plants corresponding to the unknown paint sample and a cross-correlation library searching algorithm to identify IR spectra most like the unknown in the subset of spectra identified by the search prefilters. This approach has been shown to be successful if information about the automotive manufacturer of the OEM clear coat is provided.
To obtain information about the vehicle manufacturer from an IR spectrum of an automotive clear coat, deep learning as implemented using a four-layer artificial neural network has been investigated. Specifically, the authors sought to determine the vehicle manufacturer (e.g., General Motors, Chrysler, Ford, Toyota, Nissan, and Honda) from the IR spectrum of an automotive clear coat. It has been purported that deep learning requires less data preprocessing. Raw data can (in principle) be pipelined directly to the neural network allowing it to learn patterns directly from the data for successful recognition. Selecting wavenumbers of interest or removing wavenumber regions, for example, is (in principle) not necessary as the neural network allows the data to identify key relationships within the data that are crucial for a successful classification. As shown by the results of this investigation, raw data should not be pipelined directly to the neural network, and removing wavenumber regions from the spectra can improve the performance of the model.
Downloads
Similar Publications
- Ecologically Relevant Thermal Performances of Immature Cochliomyia macellaria (Fabricius) (Diptera: Calliphoridae) and Associated Bacteria
- Refined, regionally-specific data standards reveal heterogeneity in Hispanic death records
- Approaches to mitochondrial genome sequencing using the oxford nanopore minion device