Automatic X-ray

AI Inspection for Wheels

The Artificial Intelligence ADR system is designed on the basis of a qualitative image model. The model assumes that the image is defect free and is composed from areas of relative smooth density separated by narrow regions of rapid variations (edges). Edges are not sharp structures in X-ray imagery because of the penetration and divergent rays of the central projection. In the edge regions the noise level is high. The defects are described as local deviations from the smooth intensity given by the continuity of the surface. Gas inclusions or less dense material give rise to local increase in intensity as less X-ray is absorbed. Prior to the development of digital detectors, which have high dynamic range, defects could not be repeatably detected near the edge regions due to high relative noise.

The analysis involved in the identification of defects in complex X-ray imagery takes the following steps:

1. Edge Detection

The actual location of edges provides information about the orientation of the object, and is used for the trained neural network to estimate the location of the ROI’s that the given geometry contains. Further a segmented edge image is used to mask out the edge regions in order to eliminate false detections in these regions.

2. Regions of Interest Network

The polygonal ROI’s are determined with the edge image as an input; the model is fully imperial and trained by example. Examples are derived by the operator drawing typical configurations in the teach-in phase. The ROI network outputs a mask in which the areas outside the ROI’s are masked out, and so are areas where edge may enter a ROI.

3. Adoptive Reference Subtraction

Within each ROI a fixed weight neural network estimates a smooth average intensity surface and subtracts the surface model from the actual noisy surface. The network operates with single sensitivity parameter that is calibrated in the teach mode. The parameter determines the flexibility of the adoptive surface.

4. Region of Interest Masking

The ROI originating from the neural network is used to mask out non-critical areas. Regions of non-interest (RONI) can be specified in ROI’s to minimize false reject sources.

5. Defect Identification

The final imaging process step is to convert the noisy residual resulting from the subtraction in the prior step into a two level image, which is high on the defects and low on the noisy background. A novel non-linear filter based on the so-called Hopfield-Tank neural network is employed to perform the critical task. The algorithm is unique in combining both local smoothing as a low-pass filter and the sensitivity to small defects of a high-pass filter.

6. Quality Estimation

The ROI specific configuration of defects are quantified and fed to the final classification system. The classification is based on look-up tables or user standards.

Note: Prior to the development of digital detectors, edges produced unwanted noise due to low dynamic range which is inherent in Image Intensifiers.