The arrival of low cost CMOS image sensors has opened up the market for portable multimedia applications. But while the latest standard CMOS technology allows more processing and data handling to be integrated onto a single chip, this integration is still some way off for some practical system design reasons. Nick Flaherty reports.

A typical image sensor is made up of the array of pixel diodes with the associated control circuitry, an analogue to digital converter, amplifier and enough logic to place the pixel data onto a bus. The data is then sent to a pre-processor to correct the distortions from diodes such as white balance and colour correction. Whether this processor is on the same chip as the sensor, in the camera module or on the motherboard of the mobile phone is a key question.
"The design issue is cost - what is the most cost optimised partitioning?" said Cyrille Claustres, European marketing manager for the custom solutions division of National Semiconductor. "This is why the partitioning discussion is very interesting. The sensor typically comes in a package with a window and compared to standard IC packaging as more expensive to if you bring a lot more processing into the package it could be higher cost. It depends on how the cost of the package moves. There are different lens sizes and heights of modules and these all affect the cost of the package and the ability to integrate.
"Today the module is a sensor and simple pre-processing, such as JPEG still image compression with just enough memory, a few tens of kilobytes. Then this pre-processor transfers the data to the baseband processor in the handset. This means that the baseband needs to do some work, and the amount of processing spare depends on which other services are already embedded in the handset."
As the process technology moves to smaller geometries, the switching circuitry around the pixel reduces in size compared to the pixel sensor. The choice then is to increase the sensitivity of the pixel by increasing the area, increase the number of pixels in the array to give a higher resolution, or reduce the overall size of the device, and therefore the cost, with the same sensitivity and resolution.
Then there is the trade off with the lens. The lens has to cover the sensor area, and that means using a particular focal length, which is typically 0.25in (6mm) for a VGA resolution sensor, and 0.5in (12mm) for a megapixel sensor. This focal length determines the height of the module, and often the thickness of the mobile phone.
So moving to a smaller sensor on 0.18µm gives a thinner module, except that the trend is to increase the resolution, driving the height up again.
The tradeoff also includes the material for the lens. Using a single element plastic lens is cheaper than a four element glass lens, but the difference in quality is noticeable as the absorption profile of the plastic lens is worse than the glass. But that only becomes noticeable at higher resolution.
This is why some semiconductor providers include lenses in the development kit for the sensor for engineers to evaluate the options in the optics.
Another problem is that 0.18µm is about as far as the industry can go as a process for building the sensors, says Philippe Quinio, Director of Marketing & Product Strategy, for the Imaging Division of ST Microelectronics. ST is moving to 0.18µm later in the year for its sensors.
ST is following more of a two processor partitioning for its sensors for mobile phone cameras. It has sensors with a separate processor for correcting for artifacts and handling the compression, which he sees as being used in mid to low end phones over the next two years.
Another problem with the single chip or module is that it can suffer from electromagnetic interference (EMI) problems from the radio stage of the phone on the flexible cable that links the module to the motherboard of the phone.
ST's two processor approach also gets around this by using a low voltage differential swing (LVDS) signalling protocol to connect the sensor to the processor that is inherently immune to the noise.
Motorola's Semiconductor division is also taking the two chip approach. While it has been making custom image sensors for the past five years, it has now developed a mass market device that it has included in a reference platform for a mobile phone. This allows the whole system to be optimised, not just the sensor or the module.
The VGA sensor is developed on a specialist process Motorola calls iMOS, which is moving from 0.35µm to 0.18µm early next year. The sensor is relatively simple, and uses a separate applications processor inside the module to handle the colour and defect correction and white balance. Meanwhile the module uses a three element, plastic lens to get the best colour.
The output of the module is colour corrected YUV so that the baseband processor only has to reconstruct the image and display it. This is made easier by having the module memory mapped into the baseband architecture so that the image appears as a section of memory. This also uses the memory interface running at 26MHz rather than the general purpose interface on the baseband which typically runs at 10MHz, making the image transfer significantly faster and reducing another processing load. It is this architectural optimisation that can significantly reduce the overall development time of the phone.
Atmel of the US launched its first CMOS imaging sensor alongside its existing range of megapixel CCD sensors. The Eye-On-Si range is aimed specifically at mobile phone and PDA applications with a range of resolutions planned from CIF and VGA to megapixel sensors.
National Semiconductor has just licensed imaging technology from a US start-up company called Foveon. The company developed a new image capture technology, based on CMOS, that captures all three colours - red, green and blue - simultaneously in one pixel. This is possible through an analogue design that uses three layers of photodetectors each with a different frequency sensitivity, rather than the colour filter and scan approach of traditional CMOS sensors.
However, the technology does need more complex circuits to read out the data from the sensors, but the move to 0.18µm and 90µm silicon technology reduces the relative overhead, and the pixels do not suffer from the 0.18µm limit.