Image Sensors

In a computer vision system, the image sensor converts electro-magnetic radiation (often light) into an electrical signal that represents the image. This signal is sampled and then digitised. Sensors can be single point sensors that are mechanically scanned across the image, or arrangements that are electronically scanned, such as in the TV camera. Other systems may use a single sensor and a scanned light beam, as with the laser range findcr(90].

Sensors have finite cross sectional areas. The position of the sampling point within the sensor is uncertain. Kamgar-Parsi and Kamgar-Parsi(92] have developed a model for estimating the average error due to quantisation with square sampling systems. They have extended this so that the probability o f an error being within a particular range can be obtained. It is then possible to determine whether the quantisation noise is acceptable for a particular application. Subscqucntly[93], they have extended the model to include hexagonal sampling, and from their mathematical analysis have concluded that hcxagonally shaped sensor elements yield smaller quantisation errors.

3.3.1 T V Cameras.

3.3.1.1 Raster Scanned Devices and Image Sampling.

With these devices, the 2-D image sensor is scanned a row at a time and the information transferred sequentially. In the terminology o f TV technology, the scan lines are equal to the rows o f the image array. Each line is initialised electrically by a synchronisation pulse. The set of lines that completely scan the image in one pass arc referred to as a frame. In the European standard! 1181. 625 lines

constitute a frame, and a new frame is initialised by a frame synchronisation pulse. Many cameras operate in an interlaced line scan mode in which the set of odd numbered lines, known as the odd field, are scanned first, before the even field. This facilitates lower band-width transmission.

It is possible to construct cameras with differing numbers o f lines and different scanning geometries. It is usual for the overall sensing area to be of rectangular shape with a 4:3 aspect ratio, but 1:1 ratio cameras arc available, and designs exist for programmable picture warp camcras(67]. The raster scan camera effectively initiates the sampling process as it sections the image into rows. The scanning electronics arc simplified if the line spacing is constant. All the sampling schemes considered in Section 3.2 can be sampled by a constant line spaced raster. Sampling is completed by a 1-D process, synchronised by a master clock, that divides each line into a number o f samples. This clock may increment the address to a CCD (Charged coupled device) sensor array and time the A-D (Analogue to digital) conversion process, or for a fully analogue camera, simply time the A-D converter. Convened words arc then stored or processed. An alternative camera, the line scan camera, consists o f a 1 -D array of CCD sensors. The object is moved in front of the array so that a 2-D image in built up.

For cqui-spaccd scanning o f rectangular arrays, scan lines are normally arranged to coincide with the horizontal rows of required sampling points. The value at the first point is converted at a fixed time after the line synchronisation pulse for each line. Hom|81), notes that hexagonal sampling can be realised in the same way, but that the delay between the synchronisation pulse and the first sampling point, needs to be increased by half the sampling period on alternate lines. Regular hexagonal and square geometries arc realised by calculating the sampling period so that the required horizontal physical spacing results. For hexagonal sampling, an alternative scanning scheme would be to scan parallel to the non-onhogonal y axis, as shown in Figure 3.6.

3.3.1.2 C am era Technology and Image Sampling.

In the iconoscope vacuum tube camera, light falls on a photosensitive mosaic formed by the deposition o f millions of silver globules. Each globule forms a capacitor with a common electrode provided in the construction o f the sensor. A charge is induced in each capacitor proportional to the light intensity falling on it. The charges arc then interrogated by an electron beam scanned across the mosaic target. Modem tubes have more sensitive targets than the iconoscope! 11]. The image has been sampled by the mosaic of capacitors, but the relatively wide electron beam reconstructs a

1-D analogue signal that is circularly band limited.

With solid state cameras, the sensor is a 2-D array of say CCD (Charge coupled devices). A typ­ ical array size is 512x512 elements, although a more sophisticated device may have a programmable resolution o f up to 3000x2300196] pixels. The overall sensor area is typically rectangular or square. Rows of analogue pixel values are transferred, in turn, to a line shift register, and then shifted out at a master clock rate to produce a raster scanned TV signal to a particular standard. CCD array resolutions arc comparable to the resolutions required for computer vision. Two image sampling strategies exist.

In the first strategy, the line shift register (or pixel) clock can be synchronised with the A-D converter so that the analogue pixel value is digitised while it is stable. Sampling is effected by the CCD array and depends on the array geometry. Many arrays are available for rectangular and square sampling, but these geometries should not be assumed to always be the case, as it is less expensive to fabricate devices on a hexagonal grid! 120]. The Kontron ProgRcs 3000 [961 camera ovcrsamplcs the image. This allows forthc output samples to be programmed for rectangular, square or hexagonal grids. The acquisition system from light input to digital output has been studied by McClellan! 121]. He finds non-ideal characteristics which include variations in transfer function from pixel to pixel, non-linear pixel response to illumination, and the diffusion of light within the

array. He suggests a correction procedure, but it is limited to correcting linear differences in the brightness responses o f the individual pixels. At present this would appear to be the only correction procedure available for applications operating at the video rate.

In the second strategy, the CCD array pixel clock and the A-D converter arc not synchronised. The signal is resampled at the converter. Errors will occur as the camera output signal will not be perfectly reconstructed (Oakley and Cunningham! 135]). but the choice of sampling scheme is independent of the array geometry.

For CCD sensors the output signals cannot be considered to be circularly band limited as the array elements arc often square or rectangular in shape. However, errors will be introduced if they are considered to provide a complete square or rectangular tiling of the image plane as gaps often exist between elements, and the response to illumination o f different areas within the element may be uneven.