Anomaly Detection Methods For Extraction Of Unexploded Ordnances On Aerial RGB Images

Unplanned explosion of the ammunition storage depot in Padjene, Croatia, triggered by the forest fire, happened in September 2011. Croatian research teams engaged in EU FP7 TIRAMISU project, CROMAC – CTDT Ltd. (Croatian Mine Action Center - Centre for Testing, Development and Training) and Faculty of Geodesy, University of Zagreb, promptly reacted by deployment of data acquisition module of Advanced Intelligence Decision Support System (Bajic 2010) for the aerial survey of impacted area. Troubling remnants of the explosion were scattered unexploded ordnances found in the various forms: undamaged, lightly or heavily deformed, burned or corroded. Remaining unexploded ordnances were not varying only in the form, they varied in the size also, ranging from rifle ammunition to cluster bombs. For dealing with these situations ground teams of demining experts are activated for clearance and recovery tasks. Primary objective of aerial survey in visible spectrum is to provide ground teams overview of affected area which serves as input data for planning of terrain clearance actions. This paper explores the capabilities for implementation of anomaly detection methods for extraction of unexploded ordnances on aerial RGB images acquired by consumer grade digital camera. There are two approaches in image analysis for the needs of object detection and extraction: classification and target detection. In the classification approach, the aim is to assign every pixel to the corresponding thematic class. Target detection approach, contrarily, scans the image with the aim to identify the presence of specific object or material (target). Although target detection can be regarded as binary classification with two classes: background and target, due to the difference in the essence of these approaches they are being distinguished. Number of classes is not the only difference between these approaches, term target also states that the number of the target pixels needs to be negligible compared to the number of the background class pixels. Result of this requirement is that different quality assessment methods need to be implemented depending on the approach used. Classification approach, mostly based on Bayes criterion, tends to minimize the probability of detection errors (false alarms and misses). This can easily be achieved by classifying every pixel as background, because of very sparse population of target pixels. Popular Neyman-Pearson criterion (Kay 1998), which maximizes the probability of detection while keeping the probability of false alarm at the constant rate, is more optimal for the target detection. Anomaly detection methods are a subset of target detection methods in which no a-priori information about the target spectra is available. Reed-Xiaoli (RX) (Reed & Yu 1990) algorithm is the benchmark anomaly detector: it calculates the Mahalanobis distance (Mahalanobis 1936) between the pixel under test and the background. Mahalanobis distance uses covariance matrix and mean value which can be calculated from local background (vicinity of pixel under test) or global scene covariance matrix and mean, so we distinguish local and global RX algorithm. In this paper, local and global RX algorithms were applied to the images of exploded ammunition depot acquired by Nikon D90 camera mounted on helicopter. As no ground truth data was feasible, images were visually interpreted. Object were vectorized using the object-based image analysis to reduce the human error in the manual vectorization. Experimental receiver operating curves (Kay 1998) were then constructed for both anomaly detection algorithms used. Results showed that application of anomaly detection methods can speed up the process of visual interpretation, but it cannot be used as completely automatic method for extraction of unexploded ordnances. REFERENCES Bajic, M. (2010): The Advanced Intelligence Decision Support System for the Assessment of Mine-suspected Areas, The Journal of ERW and Mine Action, 14. Kay, S. M. (1998): Fundamentals of Statistical Signal Processing, Volume 2: Detection Theory. Signal Processing, Prentice-Hall, Upper Saddle River, NJ, USA Mahalanobis, P. C. (1936): On the generalised distance in statistics, Proceedings of the National Institute of Sciences of India, 2, 49–55. Reed, I. S., & Yu, X. (1990): Adaptive multiple-band CFAR detection of an optical pattern with unknown spectral distribution, IEEE Transactions on Acoustics, Speech, and Signal Processing, 38, 1760–1770.