The use of image-based automatic defect recognition (ADR) systems in a production line often requires strict processing-time specifications. On the other hand, the typical high-performance requirement of such system calls for the use of sophisticated, computationally-complex algorithms. Addressing the conflicting requirements of fast throughput and high detection performance is a significant challenge. In this paper we present a 3D learning-based ADR approach for industrial parts. The proposed method first extracts defect candidate regions using morphological closing and template matching. Then a local registration-based approach is utilized to produce accurate defect segmentation mask. Finally, 29 features including geometric features and texture features derived from grey level co-occurrence matrix are calculated for each candidate region, and a fast random forests classifier is used to classify the candidate regions as defect or defect-free. This approach was developed into a fully automated system for detecting casting defects in aluminum industrial parts depicted in 3D Computed Tomographic (CT) images. The system was tested on 31 images with 49 cavities and porosities defects, achieving a sensitivity of 94% with an average 3.5 false detections per part.

In machine vision applications it is often desirable to make metric measurements with a single camera. However, operational constraints can limit the use of off-line calibration methods and the lack of reference planes can prevent the use of current single-view metrology techniques. In this paper, we describe the design, implementation, and validation of a visual inspection system for the measurement of angles from a single, uncalibrated camera. We provide a step-by-step, single-view metrology approach which we demonstrate in an operational setting. We first assess the imaging conditions present using hypothesis testing based on a norm-induced metric between possible world to image mappings. Our metric provides a measure of localization error derived from physical measures in the world. The hypothesis testing demonstrates that affine viewing conditions are present in our setting and we show how the affine approximation can be used to make measurements of points off the plane of interest. We employ our algorithm in a manufacturing plant to make blade angle measurements of variable stator vanes in jet engines.

Computer vision algorithms are seeing increased use in industrial inspection applications. Here, we present an "aid to visual" system that can detect post deformations of less than 0.005 inches in jet engine high pressure turbine disks. We create a gold-standard reference post from the posts of sample turbine disks and then use registration, edge detection, and curve-similarity algorithms to identify unacceptable post deformations. We address the challenges associated with adapting academic algorithms for use in functioning inspection systems. We present novel solutions to deal with practical issues such as accuracy, speed, robustness, and ease of use. Finally, we have integrated our algorithms into the software used in the work-flow of plant-factory inspectors, who will be able to identify minute post deformations that were previously difficult to detect

Many classification algorithms (such as Adaboost) aim to minimize training error, that is, the ratio of the number of mis-classified examples to the total number of examples. However, it is often desirable to minimize a more general cost metric, where distinct examples have different costs. For example, in industrial inspection, the cost of clearing a defective part is substantially greater than the cost of incorrectly alarming on a good part. Further, defects themselves vary in severity and, hence, cost. Extending boosting algorithms to account for this variable cost (cost-sensitive boosting) has been attempted by a number of investigators, resulting in a number of Adaboost variants, including AdaCost, Asymboost, CSB0, CSB1, and CSB2. These algorithms all focus only on the distribution update step in Adaboost-leaving the error computation step unchanged (or only indirectly modified). In this paper, we present a novel algorithm, termed "distributed cost boosting" (DCB), that incorporates the cost metric into the error computation step, rather than merely adjusting the distribution update step. DCB produces large empirical reductions in total cost when compared to current state-ofthe- art cost-sensitive Adaboost variants. Further, we show that distributed cost boosting provides error bounds on misclassification cost similar to AdaBoost’s bounds on training error rate.

A common difficulty encountered in tracking applications is how to track an object that becomes totally occluded, possibly for a significant period of time. Another problem is how to associate objects, or tracklets, across non-overlapping cameras, or between observations of a moving sensor that switches fields of regard. A third problem is how to update appearance models for tracked objects over time. As opposed to using a comprehensive multi-object tracker that must simultaneously deal with these tracking challenges, we present a novel, modular framework that handles each of these problems in a unified manner by the initialization, tracking, and linking of high-confidence tracklets. In this track/suspend/match paradigm, we first analyze the scene to identify areas where tracked objects are likely to become occluded. Tracking is then suspended on occluded objects and re-initiated when they emerge from behind the occlusion. We then associate, or match, suspended tracklets with the new tracklets using full kinematic models for object motion and Gibbsian distributions for object appearance in order to complete the track through the occlusion. Sensor gaps are handled in a similar manner, where tracking is suspended when the sensor looks away and then re-initiated when the sensor returns. Changes in object appearance and orientation during tracking are also seamlessly handled in this framework. Tracklets with low lock scores are terminated. Tracking then resumes on untracked movers with corresponding updated appearance models. These new tracklets are then linked back to the terminated ones as appropriate. Fully automatic tracking results from a moving sensor are presented.

We present a complete set of perceptual observables that provides a unified image description for grouping, figure-ground separation, and texture analysis. Although much progress has been made recently in treating contours and texture simultaneously for image segmentation and grouping, current approaches rely on different models for contours, regions, and texture such as one-dimensional intensity discontinuities for contours and filter bank responses for texture. This results in expensive computation that arbitrates between these disparate representations at each pixel. In our approach, salient image content such as contours, regions, and texture are represented in a common, low-level framework of image observables. We model the image as a partition of surfaces bounded by intensity discontinuities and derive perceptual measures as relations between neighboring surfaces. This enables us to extend the traditional Gestalt measures based on local edge geometry and contrast to region-based measures that jointly exploit large scale image topology, photometry, and geometry. These measures provide a natural basis for grouping on multidimensional similarity criteria and texture is directly derived as relational properties on local region neighborhoods. The viability of our model is demonstrated by applying the common observables to texture recognition, figure-ground separation, and generic image segmentation.

Human speech is inherently multi-modal, consisting of both audio and visual components. Recently researchers have shown that the incorporation of information about the position of the lips into acoustic speech recognisers enables robust recognition of noisy speech. In the case of Hidden Markov Model-recognition, we show that this happens because the visual signal stabilises the alignment of states. It is also shown, that unadorned lips, both the inner and outer contours, can be robustly tracked in real time on general-purpose workstations. To accomplish this, efficient algorithms are employed which contain three key components: shape models, motion models, and focused colour feature detectors-all of which are learnt from examples.