【正文】 high speed imaging applications and enables new imaging enhancement capabilities such as multiple capture for increasing the sensor dynamic range [4]. In this scheme, multiple images are captured at different times within the normal exposure time. Shorter exposure time images capture brighter areas of the scene,while longer exposure time images capture darker areas of the scene. The images are then bined into a single high dynamic range image. In this paper we propose to use this multiple capture capability to simultaneously form a high dynamic range image and reduce or eliminate motion blur. Our algorithm operates pletely locally – each pixel’s final value is puted using only its captured values. Moreover, our method can operate recursively, requiring the storage of only a constant number of values per pixel independent of the number of images modest putational and storage requirements make it feasible to integrate all the processing and memory needed with the image sensor on the same CMOS chip [5]. In the next section we briefly describe the image sensor pixel operation and statistical model. In Section 3 we present our algorithm. Simulation results are presented in Section 4. 2. PIXEL OPERATION AND MODELThe area image sensor used in an analog or digital camera consists of a 2D array of pixels. During capture each pixel converts incident light into photocurrent (t), for 0 t T,where T is the exposure time. This process is quite linear, and thus (t) is a good measure of incident light intensity. Since the photocurrent is too small to measure directly, it is integrated onto a capacitor and the charge Q(T) (or voltage) is read out at the end of exposure time T. Dark current and additive noise corrupt the output signal charge. The noise can be expressed as the sum of three independent ponents, (i)shot noise U(T)~N(0。 q), where q is the electron charge, (ii) readout circuit noise V (T) (including quantization noise) with zero mean and variance , and (iii)reset noise C ~N(0,) caused by resetting the capacitor prior to capture. Thus the output charge from a pixel can be expressed as (1)provided Q(T) Q sat,the saturation charge, also referred to as well capacity. If photocurrent is constant over exposure time, SNR can be expressed as (2) Note that SNR increases with , first at 20dB per decade when reset and readout noise variance dominates, and then at 10dB per decade when shot noise variance dominates. SNR also increases with T. Thus it is always preferred to have the longest possible exposure time. Saturation and change in photocurrent due to motion, however, makes it impractical to make exposure time too long. We illustrate the effect of saturation and motion and how multiple capture may mitigate their effects via the examples in Figures 1 and 2. The first plot in Figure 1 represents the case of a constant low light, where photocurrent can be well estimated from Q(T). The second plot represents the case of a constant high light, where Q(T) = Qsat and the photocurrent cannot be well estimated from Q(T). The third plot is for the case when light changes during exposure time, ., due to this case, photocurrent at the beginning of exposure time(0) again cannot be well estimated from Q(T). To avoid saturation and the change of (t) due to motion,exposure time may be shortened, ., to in Figure in conventional sensor operation, exposure time is set globally for all pixels, this results in reduction of SNR, especially for pixels with low light. This point is further demonstrated by the images in Figure 2, where a bright square object moves diagonally across a dark background. If exposure time is set long to achieve high SNR, it results in significant motion blur as shown in image (b) of Figure 2. On the other hand if exposure time is set short, SNR deteriorates resulting in the noisy image (c) of Figure 2. Recent advances in CMOS image sensor technology makes it possible to capture and nondestructively read out, ., without reset, multiple images within a normal exposure time [4].Using this multiple capture capability one can in effect adapt the pixel exposure time to the lighting condition. For the examples in Figure 1, if we capture four images at , 2 , 3 ,and T = 4 , the photocurrent for the high light pixel can be estimated using the images captured at  and 2 , while for the low light pixel can be estimated using the four images. Motion blur in the third case can be reduced by using the four captures to estimate photocurrent at the beginning of exposure time (0). In the following section we derive an optimal recursive pixel photocurrent estimator from multiple captures and show how motion blur can be detected and reduced. Image (d) in Figure 2 of the moving square object is produced using four captures and the algorithm described in the next section.Figure 2: (a) Ideal image. (b) Long exposure time image. (c)Short exposure time image. (d) Image produced by applying our algorithm to 4 captures. 3. IMAGE FORMATION AND BLUR RESTORATION Our image formation and motion blur restoration algorithm operates on n image captures at 。 2。 : : : 。 n = T as follows:1. Capture first image, set k = 1.2. For each pixel: Use the current estimation algorithm to find the photocurrent estimate from Q( ).3. Capture next image.4. For each pixel: Use the motion detection algorithm to check if motion/saturation has occurred i. Motion detected: Set final photocurrent estimate = . ii. No Motion detected or decision deferred: Use the current estimation algorithm to find from Q((k + 1) ) and and set k = k + 1.5. Repeat steps 3 and 4 until k = n.In the following subsection we describe a recursive algorithm for estimating photocurrent, and in subsection we describe a heuristic algorithm for performing motion detection.. Photocurrent EstimationWe simplify