Image acquisition
A Software framework has been developed to allow images to be acquired and processed and different algorithms tested.

Prior to conducting real PCB inspection, a golden image is captured and stored in a computer which serves as a reference. In this golden image, all components exist on the PCB with sufficient precision according to manufacturing specifications. Subsequently, live images of a PCB being inspected in a production line are taken and subtracted from the golden image. The resultant residual image reveals any differences between the target PCB image and the reference image. Where there are bright pixels remaining, a component is likely to be missing. This can be done automatically by a whole image field intensity check, thus achieving an automated inspection of missing components.

Pre-processing
Due to the limited field of view of the imaging system with a sufficient image resolution, it is always necessary to take several local images of a PCB. These local images are then stitched together to form a complete image. With a proper mechanical positioning system, this can be implemented automatically. The only potential problem is misalignment between different local images as shown in figure 4 because of limited precision of the mechanical positioning system.

Image Registration
To correct for any mechanical misalignments and hence different X and Y pixel displacements of the PCB in the golden reference image and the successive target images image registration is performed to compensate for any differences.

PCB manufacturers normally put fiducial marks in the printed circuit artwork to provide common measurements for all steps in the assembly process. Fiducial marks are often put locally to particular components that require more precise location. These fiducial marks can be used to guide the placement of corresponding local image to the correct position. In our AOI prototype for each image the user is required to indicate to the registration algorithm where the fiducial mark is. Ideally, at least two fiducial marks should be available in each local image to facilitate registration.

Image Subtraction
Given two images, one golden image and a live image, it is digitally convenient to perform an image subtraction. To enable the subtracted image to be visualised appropriately, intensity scale adjustment is performed. This is to project the subtracted intensities to a dynamic range available for the computer display, either by linear projection, step-size projection, or non-linear projection.

It is apparent that, to ensure a successful image subtraction that reveals only those differences corresponding to missing components in a target PCB, all the imaging parameters must be kept constant, both for the golden image and for the live images of PCBs in a production line. For those PCBs that have some highly reflective surfaces, a slight change in illumination will cause considerable variation in the intensities of the captured image. As a result, significant residual intensities remain after image subtraction. This is one of the main sources of false alarming in the PCB inspection. Figure 5 shows the results of one typical PCB inspection. Most of the residual bright intensities are associated with real defects, i.e. missing components, while some stem from the shinning variations. They need to be further characterized based on context analysis in the next step.

Fig. 5a Golden Image	Fig. 5b Target Image
Fig. 5c The final residual image by subtracting (5b) from (5a).	Fig. 5d Identification of possible defects by software

Residual image analysis and defect identification
The residual intensities are analysed by taking account of the context of original images. Firstly, a threshold operation is performed to highlight the residual pixels with high intensities. The optimal threshold is found from the histogram of the original image. Those pixels with high intensities are characterised as object pixels while others are characterised as background pixels. A grouping operation is followed to merge together separated pixels belonging to a common component:- operations of medium filtering, corrosion, and void filling are applied so that isolated bright residual pixels can be eliminated and only those bright spots corresponding to real missing components are retained. Subsequently, potential defects are identified when there are a group of bright pixels and in addition where their original image does not exhibit near-saturated intensities. Near-saturated intensities are mostly likely corresponding to surface shining. In fig. 5 green boxes are overlayed on the original target image indicating suspected defects. The green arrow indicates a false alarm due to shining of the soldering surface; this is highlighted for illustration purposes only.