K.U.Leuven > ESAT > PSI > Visics > Research > Topics > Item 3.2

SOQUETEC

L. Van Gool, F. Feyaerts, P. Pollet, P. Vanroose

There are three target applications within the Soquetec project: inspection of seeds (for germination prediction), inspection of tiles (glazing defects and colour deviations) and inspection of eggs (dirt and crack defects).

Visual inspection of seeds

The seeds were measured with an experimental set-up using an Agfa RGB flatbed scanner with 12-bit colour depth and a spatial resolution of 2400 dpi.

Rotation and translation invariant parameterisation of the contour was first calculated. Next, curvature of this contour, ratio of contour length and enclosed area, FFT and rotation and translation invariant spatial moments were calculated. Possible separation between good and bad seeds is estimated as shown in the following table.

Visual inspection of tiles

Quality inspection on tiles includes two separate tasks: identification of surface defects and identification of colour and pattern defects. Identification of surface defects requires a monochrome camera with high resolution in combination with a special illumination and camera set-up. Mirror reflection highlights the defects between the light stripes and suppresses the reflection inside the light stripes. For the identification of colour and pattern a colour camera and diffuse white illumination was used.

After tile detection, the projective deformation of the tile image was compensated enabling pixel-wise pattern matching. The difference image between the reference for the best fitted colour batch enabled to detect pattern defects and some large (camera resolution) surface defects.

The image on the left shows a tile after compensation for projective deformations. The image on the right shows the difference image indicating defective regions.

Visual inspection of eggs

Software was developed to process image regions (ROI's) rather than entire images to reduce the processing time. After the eggs were approximated by an ellipse, the observed intensity variation caused by the 3D shape of the eggs was corrected based on polynomial fitting (second order) through a set of observed intensities on the egg shape. The corrected intensity, i.e. the average intensity increased with the observed error with respect to the calculated polynomial, eliminated only polynomial variation of the intensity profile (between the edges of the ellipse). Specular reflections of the light tubes on "glazed eggs" were left intact. An image reconstruction based on (the least significant) principal components could be used to reduce not only the illumination variation but also all other natural variation of egg shape, colour, ... However, due to the large variation on those parameters (size, colour, glazing and position of the eggs in the images), this method appeared not feasible.

Features were generated for segmented regions inside the egg (fast and segmented regions outside the egg are irrelevant, i.e., background) after elimination of regions of specular reflection of the light tubes and egg-background. All the features were rotation and translation invariant, i.e., the features did not change with location and orientation of the egg. The feature consisted of combined rotation and translation invariant spatial and photometrical moments up to order 6: x^ay^bI^c (a+b+c ≤ 6), rotation and translation invariant description of position of the segmented regions inside the egg and rotation and translation invariant contour description of the segmented regions (measure of compactness for the region: total squared error of the distances between the edge points and the fitted ellipse and ratio of the contour length and the enclosed area).

Ellipse fitting	Intensity correction	Segmentation