Donald Baxter

SXGA pinned photodiode CMOS image sensor in 0.35 /spl mu/m technology

A 30 frames/s SXGA 5.6 /spl mu/m pinned photodiode pixel column parallel CMOS image sensor achieves 340 /spl mu/V noise floor and 40 pA/cm/sup 2/ dark current. Performance is limited by pixel 1/f noise, not by the ADC noise floor of 140 /spl mu/V. The column ADC memory employs a custom DRAM to save area. The sensor utilizes a 0.35 /spl mu/m 1P 3M CMOS process.

Source follower noise limitations in CMOS active pixel sensors

CMOS imagers are commonly employing pinned photodiode pixels and true correlated double sampling to eliminate kTC noise and achieve low noise performance. Low noise performance also depends on optimisation of the readout circuitry. This paper investigates the effect of the pixel source follower transistor on the overall noise performance through several characterization methods. The characterization methods are described, and experimental results are detailed. It is shown that the source follower noise can be the limiting factor of the image sensor and requires optimisation.

Development of the I3A CPIQ spatial metrics

The I3A Camera Phone Image Quality (CPIQ) initiative aims to provide a consumer-oriented overall image quality metric for mobile phone cameras. In order to achieve this goal, a set of subjectively correlated image quality metrics has been developed. This paper describes the development of a specific group within this set of metrics, the spatial metrics. Contained in this group are the edge acutance, visual noise and texture acutance metrics. A common feature is that they are all dependent on the spatial content of the specific scene being analyzed. Therefore, the measurement results of the metrics are weighted by a contrast sensitivity function (CSF) and, thus, the conditions under which a particular image is viewed must be specified. This leads to the establishment of a common framework consisting of three components shared by all spatial metrics. First, the RGB image is transformed to a color opponent space, separating the luminance channel from two chrominance channels. Second, associated with this color space are three contrast sensitivity functions for each individual opponent channel. Finally, the specific viewing conditions, comprising both digital displays as well as printouts, are supported through two distinct MTFs.

A line-based HDR sensor simulator for motion artifact prediction

Modeling only a HDR’s camera’s lens blur, noise and sensitivity is not sufficient to predict image quality. For a fuller prediction, motion blur/artifacts must be included. Automotive applications are particularly challenging for HDR motion artifacts. This paper extends a classic camera noise model to simulate motion artifacts. The motivation is to predict, visualize and evaluate the motion/lighting flicker artifacts for different image sensor readout architectures. The proposed motion artifact HDR simulator has 3 main components; a dynamic image source, a simple lens model and a line based image sensor model. The line based nature of image sensor provides an accurate simulation of how different readout strategies sample movement or flickering lights in a given scene. Two simulation studies illustrating the model’s performance are presented. The first simulation compares the motion artifacts of frame sequential and line interleaved HDR readout while the second study compares the motion blur of an 8MP 1.4μm, 5MP 1.75μm and 3MP 2.2μm image sensors under the same illumination level. Good alignment is obtained between the expected and simulated results.

The subjective importance of noise spectral content

This paper presents secondary Standard Quality Scale (SQS2) rankings in overall quality JNDs for a subjective analysis of the 3 axes of noise, amplitude, spectral content, and noise type, based on the ISO 20462 softcopy ruler protocol. For the initial pilot study, a Python noise simulation model was created to generate the matrix of noise masks for the softcopy ruler base images with different levels of noise, different low pass filter noise bandwidths and different band pass filter center frequencies, and 3 different types of noise: luma only, chroma only, and luma and chroma combined. Based on the lessons learned, the full subjective experiment, involving 27 observers from Google, NVIDIA and STMicroelectronics was modified to incorporate a wider set of base image scenes, and the removal of band pass filtered noise masks to ease observer fatigue. Good correlation was observed with the Aptina subjective noise study. The absence of tone mapping in the noise simulation model visibly reduced the contrast at high levels of noise, due to the clipping of the high levels of noise near black and white. Under the 34-inch viewing distance, no significant difference was found between the luma only noise masks and the combined luma and chroma noise masks. This was not the intuitive expectation. Two of the base images with large uniform areas, ‘restaurant’ and ‘no parking’, were found to be consistently more sensitive to noise than the texture rich scenes. Two key conclusions are (1) there are fundamentally different sensitivities to noise on a flat patch versus noise in real images and (2) magnification of an image accentuates visual noise in a way that is non-representative of typical noise reduction algorithms generating the same output frequency. Analysis of our experimental noise masks applied to a synthetic Macbeth ColorChecker Chart confirmed the color-dependent nature of the visibility of luma and chroma noise.

Embedded Signal Approach to Image Texture Reproduction Analysis

Since image processing aimed at reducing image noise can also remove important texture, standard methods for evaluating the capture and retention of image texture are currently being developed. Concurrently, the evolution of the intelligence and performance of camera noise-reduction (NR) algorithms poses a challenge for these protocols. Many NR algorithms are ‘content-aware’, which can lead to different levels of NR being applied to various regions within the same digital image. We review the requirements for improved texture measurement. The challenge is to evaluate image signal (texture) content without having a test signal interfere with the processing of the natural scene. We describe an approach to texture reproduction analysis that uses embedded periodic test signals within image texture regions. We describe a target that uses natural image texture combined with a multi-frequency periodic signal. This low-amplitude signal region is embedded in the texture image. Two approaches for embedding periodic test signals in image texture are described. The stacked sine-wave method uses a single combined, or stacked, region with several frequency components. The second method uses a low-amplitude version of the IEC-61146-1 sine-wave multi-burst chart, combined with image texture. A 3x3 grid of smaller regions, each with a single frequency, constitutes the test target. Both methods were evaluated using a simulated digital camera capture-path that included detector noise and optical MTF, for a range of camera exposure/ISO settings. Two types of image texture were used with the method, natural grass and a computed ‘dead-leaves’ region composed of random circles. The embedded-signal methods tested for accuracy with respect to image noise over a wide range of levels, and then further in an evaluation of an adaptive noise-reduction image processing.

Calibration and adaptation of ISO visual noise for I3A's Camera Phone Image Quality initiative

The I3A Camera Phone Image Quality (CPIQ) visual noise metric described is a core image quality attribute of the wider I3A CPIQ consumer orientated, camera image quality score. This paper describes the selection of a suitable noise metric, the adaptation of the chosen ISO 15739 visual noise protocol for the challenges posed by cell phone cameras and the mapping of the adapted protocol to subjective image quality loss using a published noise study. Via a simple study, visual noise metrics are shown to discriminate between different noise frequency shapes. The optical non-uniformities prevalent in cell phone cameras and higher noise levels pose significant challenges to the ISO 15739 visual noise protocol. The non-uniformities are addressed using a frequency based high pass filter. Secondly, the data clipping at high noise levels is avoided using a Johnson and Fairchild frequency based Luminance contrast sensitivity function (CSF). The final result is a visually based noise metric calibrated in Quality Loss Just Noticeable Differences (JND) using Aptina Imagings subjectively calibrated image set.

Toward a quantitative visual noise evaluation of sensors and image processing pipes

The evaluation of sensors performance in terms of signal-to-noise ratio (SNR) is a big challenge for both camera phone manufacturers and customers. The first ones want to predict and assess the performance of their pixel while the seconds require being able to benchmark raw sensors and processing pipes. The Reference SNR metric is very sensitive to crosstalk whereas for low-light issue, the weight of sensitivity should be increased. To evaluate noise on final image, the analytical calculation of SNR on luminance channel has been performed by taking into account noise correlation due to the processing pipe. However, this luminance noise does not match the perception of human eye which is also sensitive to chromatic noise. Alternative metrics have been investigated to find a visual noise metric closer to the human visual system. They have been computed on five pixel technologies nodes with different sensor resolutions and viewing conditions.