An Automated Vision Based Flaw Detection System for Vehicle Inspection
We are developing a computer vision system for Creative Video Associates Limited (CVA) which is capable of detecting flaws, such as scratches, in vehicle bodywork. The system is intended to analyse video images of stationary vehicles to determine whether there has been any damage to the body subsequent to manufacture. Preliminary discussions with CVA have suggested that techniques which we have already developed at the University of Bristol could be adapted for this task.In earlier work we have developed a neural-network based vision system, implemented in software on a standard graphics workstation, which can automatically divide a natural scene into regions corresponding to objects and correctly classify them, to high accuracy, across a wide range of object types (i.e. car bodies, windows, roads, etc.).
The motor industry utilises video cameras to obtain images of vehicles, prior to shipment, for the purpose of establishing that the vehicle was fault free when leaving the site. At present these are simply stored as images to be looked at in the event of a fault being detected subsequently. It would clearly be beneficial if an automatic system could analyse the videos in real time and report any flaws detected.
Our approach to this problem is to use a neural network classifier whose inputs are features extracted from video images of parts of the vehicle and whose output will indicate the presence of surface imperfections. The advantage of a neural network approach is that it can be trained on a set of sample images of both perfect vehicles and those containing the characteristic flaws which we will be seeking to detect. This avoids the need for a predefined model of the vehicle and will allow the system to operate over a range of vehicle types. A further advantage is that separate networks can be trained, if necessary, to deal with varying conditions eg. cars fresh from the assembly line versus those which have been in storage lots for a period of time. Separate networks can also be specifically trained to detect imperfections in awkward positions such as close to edges of doors etc. The general advantage, therefore, of a neural-network approach is its ability to generalise, giving flexibility over a range of similar recognition tasks.
To simplify the task for the network classifier, we intend to preprocess the images to remove most of the common features that might be confused with imperfections (door edges and handles, window frames, emblems etc.). This will be achieved by building a visual model of the car, as seem from each camera, which can be aligned with the image and used to identify expected features.
Staff and Students
Barry Thomas, Tom Troscianko, Michael Evans, Chris SetchellPublications
- MacKeown, W.P., P. Greenway, B.T. Thomas and W.A. Wright, (1994), ``Contextual Image Labelling with a Neural Network'', IEE Vision, Image and Signal Processing, 141,4, pp. 238-244
- McCloskey, G., W.P. Mackeown and B.T. Thomas, (1995), ``Dimensionality Reduction of Feature Vectors for Classification'', Irish Neural Network Conference, Maynooth, Sept 1995.
- Troscianko, T.,N.W. Campbell, W.P.J. Mackeown and B.T. Thomas, (1995), ``Segmentation of natural scenes by means of colour and texture", Perception 24 (supplement), 18.