Continuous improvement of product quality is crucial to the successful and competitive automotive manufacturing industry in the 21st century. The presence of surface porosity located on flat machined surfaces such as cylinder heads/blocks and transmission cases may allow leaks of coolant, oil, or combustion gas between critical mating surfaces, thus causing damage to the engine or transmission. Therefore 100% inline inspection plays an important role for improving product quality. Although the techniques of image processing and machine vision have been applied to machined surface inspection and well improved in the past 20 years, in today's automotive industry, surface porosity inspection is still done by skilled humans, which is costly, tedious, time consuming and not capable of reliably detecting small defects. In our study, an automated defect detection and classification system for flat machined surfaces has been designed and constructed. In this paper, the importance of the illuminant direction in a machine vision system was first emphasized and then the surface defect inspection system under multiple directional illuminations was designed and constructed. After that, image processing algorithms were developed to realize 5 types of 2D or 3D surface defects (pore, 2D blemish, residue dirt, scratch, and gouge) detection and classification. The steps of image processing include: (1) image acquisition and contrast enhancement (2) defect segmentation and feature extraction (3) defect classification. An artificial machined surface and an actual automotive part: cylinder head surface were tested and, as a result, microscopic surface defects can be accurately detected and assigned to a surface defect class. The cycle time of this system can be sufficiently fast that implementation of 100% inline inspection is feasible. The field of view of this system is 150mm×225mm and the surfaces larger than the field of view can be stitched together in software.
It is practically impossible to collect an exhaustive set of minefield data for all different environment conditions, diurnal cycle, terrain conditions and minefield layouts. Such a data collection may in fact be even more expensive to ground truth, register and maintain than to acquire. This paper explores minefield synthesis using patch-based sampling of previously acquired airborne mid-wave infra-red (MWIR) images. The main idea is to synthesize a new (minefield) image by selecting appropriate small patches from the existing images and stitching them together in a consistent manner to simulate realistic imagery for different minefield scenarios. The selected patches include those from different background types, emplaced cultural clutter and different mine types. We assume a first order Markov model for the image so that the image-patch at a particular location is dependent on the characteristics of the image patch in the immediate neighborhood only. The proposed model is capable of generating any desired terrain condition (homogenous or inhomogeneous) based on a given terrain map. In addition, it supports generating different minefield layouts such as patterned or scattered minefields using mine patches from appropriate backgrounds. The paper presents representative synthesized minefield imagery and image sequences using previously collected real airborne data. Minefield image data synthesized using this procedure should be valuable in an airborne minefield detection program for evaluating most mine detection as well as minefield detection algorithms.
The aim of an anomaly detector is to locate spatial target locations that show significantly different spectral/spatial characteristics as compared to the background. Typical anomaly detectors can achieve a high probability of detection, however at the cost of significantly high false alarm rates. For successful minefield detection there is a need for a further processing step to identify mine-like targets and/or reject non-mine targets in order to improve the mine detection to false alarm ratio. In this paper, we discuss a number of false alarm mitigation (FAM) modalities for MWIR imagery. In particular, we investigate measures based on circularity, gray scale shape profile and reflection symmetry. The performance of these modalities is evaluated for false alarm mitigation using real airborne MWIR data at different times of the day and for different spectral bands. We also motivate a feature based clustering and discrimination scheme based on these modalities to classify similar targets. While false alarm mitigation is primarily used to reject non-mine like targets, feature based clustering can be used to select similar-looking mine-like targets. Minefield detection can subsequently proceed on each localized cluster of similar looking targets.
One of the primary lessons learned from airborne mid-wave infrared (MWIR) based mine and minefield detection research and development over the last few years has been the fact that no single algorithm or static detection architecture is able to meet mine and minefield detection performance specifications. This is true not only because of the highly varied environmental and operational conditions under which an airborne sensor is expected to perform but also due to the highly data dependent nature of sensors and algorithms employed for detection. Attempts to make the algorithms themselves more robust to varying operating conditions have only been partially successful. In this paper, we present a knowledge-based architecture to tackle this challenging problem. The detailed algorithm architecture is discussed for such a mine/minefield detection system, with a description of each functional block and data interface. This dynamic and knowledge-driven architecture will provide more robust mine and minefield detection for a highly multi-modal operating environment. The acquisition of the knowledge for this system is predominantly data driven, incorporating not only the analysis of historical airborne mine and minefield imagery data collection, but also other “all source data” that may be available such as terrain information and time of day. This “all source data” is extremely important and embodies causal information that drives the detection performance. This information is not being used by current detection architectures. Data analysis for knowledge acquisition will facilitate better understanding of the factors that affect the detection performance and will provide insight into areas for improvement for both sensors and algorithms. Important aspects of this knowledge-based architecture, its motivations and the potential gains from its implementation are discussed, and some preliminary results are presented.