Keeping a watch on the numbers

Character printing control system aides Swiss watch manufacturers

The Swiss are known for the art of keeping time. A Swiss watch is synonymous with quality. With such a strong worldwide reputation to uphold, it is hardly surprising that watch makers adhere to the strictest standards of quality.

The hundreds of parts that fit inside the watch must be machined with utmost precision for the watch to keep perfect time. Close attention must also be paid to the face, referred to as the dial. Like everything else that is manufactured for consumers, the product's acceptance in the marketplace greatly depends on its appearance.

Whatever the material, the numbers are printed on the watch dials, opening doors for possible printing defects. Spattered ink might be found as spots outside the numbers. White specks that lie within the number indicate the printed character is missing ink. Spreading ink might also produce thicker numbers than what is acceptable. These are precisely the issues that a dial manufacturer brought to Visiosoft SA in Lausanne, Switzerland.

Time for quality control

Visiosoft developed a system for an inspection process known as aesthetic character printing control (ACPC). It differs from optical character recognition (OCR) in that ACPC searches for printing defects such as spots inside or outside the character. ACPC considers such parameters as "the edge's quality, the width of the character, and even the relative position of the different characters," says Alain Favre, Business Unit Manager for Visiosoft. OCR and OCV systems actually read the font, which is not the case in an aesthetic control.

"The pride of watch makers is to exhibit beautiful fonts," explains Favre. These fonts, however, are generally not standard fonts, and since they differ considerably between different watch models, OCR does not make a suitable inspection method. The backgrounds of the dials may also slightly differ from one watch to another. Typically, controlling the backgrounds requires texture analysis, a technique which is related to gray levels co-occurrence, and differs from the analysis made in the region of the printed characters. At least two separate analyses are necessary to mask the background with respect to the characters. Consequently, an aesthetic character printing control device must be able to adapt itself to these situations. Simple pattern matching (normalized grayscale correlation) or usual Print Quality Inspection techniques are therefore unusable for such an inspection system.

The system

Visiosoft's system is comprised of a Sony XC-55 camera fitted with a Computar TEC-m55 lens, a PC running at Pentium-III speeds, and a Matrox Meteor-II /Multi-Channel frame grabber. The system uses a custom lighting solution derived from existing commercial components. The analysis is performed by the Geometric Model Finder (GMF) and Blob Analysis modules of the Matrox Imaging Library (MIL).

Watch the results

The watch inspection solution is to analyze the scene by splitting the object into smaller, manageable regions. The inspector places the dial on a special mechanical piece, and since the position of the dial is not exactly reproducible, the image analysis depends first on locating its edge (as shown in Figure 1), and then localizing the characters on the visible part of the dial. "GMF from Matrox gives good results for this process," says Favre. The image is processed and analyzed, and the dial automatically rotates to the next number, until all the dial's numbers have been inspected. The images are compared to test images; if the dial fails the inspection, the operator discards it. Since the hours on the watch's dial follow a sequence, the models for the individual number characters, (1, 2, 3...) can also be sequenced, which saves time. Furthermore, due to the sophisticated nature of the fonts, the model for "12" is a separate model, and not a combination of the models for "1" and "2"; therefore, the available parameters in the GMF must be adjusted with care. The printing process adds to the difficulty. Ink fluidity might be inconsistent and produce "fat" characters that should be considered well formed. This means a fatter or thinner 12 might be mistaken for a 1 and 2.

Applying the GMF algorithms yields a result image whose edges are optimally adjusted over the sample image (Figure 2). The edges fit the contour of the visible character. Then, a fill operation is performed on the image, which yields the best mask for the visible character. If the sample character itself contains holes, like a 0 or an 8, it is necessary to fill the remaining holes and mask them correctly. Masking the holes is necessary to separate the character from the background.

Once the characters are localized, the model of the character is used to separate the character's area from the background. The quality of each region is measured by different algorithms. Details of these algorithms depend on the nature of the material, such as iron, plastic or aluminum, any paintings and various protection layers. The measures of the perceptual uniformity vary enormously from one case to the other.

The greatest challenge in developing the system was delimiting the different parts of the surface, and developing a system that could adequately analyze and sort the defects found in each part. Visiosoft's inspection system was developed for a single customer, and has improved the efficiency of the inspection process by reducing both the control time and the amount of human intervention. When the dials were inspected by hand, the quality of the dial depended on the operator's standards. Automated inspection systems provides complete objectivity, which is an essential component in a quality control system. Visiosoft's system has proven to be a valuable tool for the watch industry. Other dial manufacturers have expressed an interest in the system, and the company is looking to modify their technology to inspect other parts that are manufactured for Switzerland's watch industry.

Figure 1 shows how the outside mask is determined: (a) original image; (b) displays the window used to find the dial's border; (c) displays the detection and hence the Y coordinate of the border; (d) the exclusion of the dark (red) part of the figure in using a correctly repositioned mask.

Figure 2 (a) displays the original image and (b) shows the superimposed model ( in red ).The "fill hole" operation applies first to the outside contour (c) and yields to (d). Applying "fill hole" to internal contours (e) yields to (f). The results of subtracting (f) from (d) optimally delimits the character.