Fixed pattern noise compensation with low memory requirements

[0001] (1) Field of the Invention

[0002] This invention relates generally to image processing and relates more particularly to a method to reduce the fixed pattern noise in an obtained image.

[0003] (2) Description of the Prior Art

[0004] CMOS-based cameras are being plagued by a defect known as “fixed pattern noise” (FPN) This is an annoying stationary background pattern in the image that results from small differences in the behaviour of the individual pixel amplifiers. FPN (also called nonuniformity) is the spatial variation in pixel output values under uniform illumination due to device and interconnect parameter variations (mismatches) across the sensor. FPN is fixed for a given sensor, but varies from sensor to sensor, so if v_o is the nominal pixel output value (at uniform illumination), and the output pixel values (excluding temporal noise) from the sensor are v_oij for 1≦i≦n and 1≦j≦m, then the fixed pattern noise is the set of values Δv_oij=v_oij−v_o.

[0005] FPN consists of offset and gain components, the offset component causes more degradation in image quality at low illumination. CMOS (Active Pixel Sensors—APS— and Passive Pixel Sensors—PPS—) suffer from column FPN, which appears as “stripes” in the image and can result in significant image degradation. CCD image sensors only suffer from pixel FPN due to spatial variation in photodetector geometry and dark current. Using CMOS technology image sensors pixel transistors cause additional pixel FPN and column amplifiers cause column FPN.

[0006] In prior art, the normal way to reduce FPN is to do a so-called Correlated Double Sampling (CDS). This involves sampling the output twice, once right after reset and the second time with the signal present. The offset and reset noises are reduced by taking the difference. Most sensors cannot implement a true CDS and thus cannot fully compensate offset FPN.

[0007] Dark frame subtraction comprises to capture one or more dark frames under the actual picture taking conditions (exposure time and sensor temperature), and the frame or average of multiple frames is subtracted from the image. Dark frame compensation can be used but requires an external shutter and a full frame memory and is therefore an expensive solution to reduce the FPN problem.

[0008] There are patents dealing with the suppression of FPN:

[0009] U.S. Pat. No. 6,061,092 to Bakhle et al. describes a method and an apparatus for dark frame cancellation for CMOS sensor-based tethered video peripherals. The elimination of dark fixed pattern noise (DFPN) for tethered CMOS sensor-based digital video cameras is supported by supplying and maintaining a host-based dark image cache. Since the camera is tethered to a host computer system such as a PC, it takes advantage of the storage and processing capabilities of the host to manage the cache. By using a dark image cache for updating of the currently applicable dark image for DFPN cancellation processing, operation of the camera shutter for acquiring dark images is dramatically reduced. Dark images are obtained at different integration, gain, and temperature operating characteristics of the camera and stored in the cache. The cached dark images are referenced on the host according to a fixed, predetermined column of data in video frames generated by the CMOS sensor image array of the camera. The dark column data represents a portion of the CMOS sensor image array that is permanently and totally shadowed for use during DFPN cancellation processing.

[0010] U.S. Pat. No. 6,215,113 to Chen et al. discloses a CMOS active pixel sensor with motion detection including a photo diode, a reset switch, two sample and hold circuits, and two readout circuits. In addition to a CMOS sensor array, a row of CMOS active pixel sensors (APS) can be added to reduce the fixed pattern noise caused by process variations. Each of the CMOS APS cells of this additional row is the same as the CMOS APS cells in said CMOS sensor array except that the photo diode area thereof is covered by a metal layer to block the incident light. Therefore, the readout signals of the pixels in this additional row should be fixed signals after the reset process and should be independent of the incident light. When an image signal of a selected APS cell is read out, a signal of the shielded APS cell in the same column with the selected APS cell is also read out. The final image is the difference of these two signals. Since these two readout values pass through the same readout circuit, the effect of process variation can be cancelled and the fixed pattern noise can be reduced.

[0011] U.S. Pat. No. 6,366,320 to Nair et al. discloses a semiconductor circuit having an analog storage array and a sense amplifier array in which each sense amp cell generates a differential signal pair in response to receiving first and second signals from the storage array. The sense amp array is a row of sense amplifier cells, one sense amp cell for each column. The sense amplifier cells implement correlated double sampling (CDS) of pixel signals. CDS is used to reduce fixed pattern noise in pixel signals. Fixed pattern noise refers to errors in otherwise identical pixel signal values that are caused by manufacturing variations between the pixels. CDS requires sampling both the pre-exposure “reset value” and the post-exposure “exposed value” of each pixel signal. The two values are then transferred as a differential signal pair to a signal-processing pipe. There the values are correlated or simply subtracted to cancel the errors caused by differences between reset values and circuitry in the pixels of the manufactured array.

[0012] A principal object of the present invention is to achieve a compensation of fixed pattern noise in digital images with low memory requirement.

[0013] A further object of the present invention is to achieve a compensation of fixed pattern noise in digital images without the need of an external shutter

[0014] In accordance with an object of this invention a further method to compensate fixed pattern noise of digital images without the need of an external shutter is achieved. Said method comprises first, providing a non-volatile memory, a processing device, calibration frames, and an image sensor. The steps of said method comprises to derive FPN data for each sensor using calibration frames, to store FPN data measured, to retrieve FPN data for each picture taken and to apply FPN data on each pixel of pictures taken to compensate fixed pattern noise.

[0015] In accordance with the objects of this invention a further method to compensate fixed pattern noise of digital images without the need of an external shutter and with low memory requirements is achieved. Said method comprises, first, providing a non-volatile memory, a processing device, calibration frames, and an image sensor. The first steps of said method comprises to derive FPN data for each sensor using calibration frames, reduce said FPN data derived, to perform compression on FPN data and to store factor used for data reduction, compressed FPN data, compression table used, and distribution measurement results in said non-volatile memory. Further steps comprise to perform decompression on FPN data, to perform scaling of reduced FPN data using current statistical distribution measurement, and to apply individual compensation value on each pixel.

[0016] In accordance with the objects of this invention a further method to compensate fixed pattern noise of digital images without the need of an external shutter and with low memory requirements is achieved. Said method comprises first, providing a non-volatile memory, a processing device, calibration frames, and an image sensor. The first steps of said method comprise to derive FPN data for each sensor using calibration frames, to quantize said FPN data derived, to generate an optimized Huffman table, to perform Huffman compression on FPN data using optimized Huffman table, and to store quantization factor, compressed FPN data, Huffman compression table used, and distribution measurement results. Further steps comprise to perform Huffman decompression on FPN data, to perform quantization scaling using current statistical distribution measurement, and to apply individual compensation value on each pixel.

[0017] In accordance with the objects of this invention a system to compensate fixed pattern noise of digital images without the need of an external shutter and with low memory requirements is achieved. Said system comprises a non-volatile memory, a processing device, and an image sensor.

[0022] The preferred embodiment disclose a system and a method to reduce the fixed pattern noise (FPN) without the need of an external shutter requiring a small memory only as compared with prior art using Correlated Double Sampling (CDS) or Dark Frame Subtraction where a full frame memory is required.

[0023] In the novel method invented the FPN compensation is based on processing done during production. The fixed pattern noise is here defined as the fixed pattern seen in the individual pixel offsets. This noise increases in amplitude with increasing temperature and exposure time.

[0024] The fixed pattern noise is uncorrelated noise but it has a statistical distribution that can be scaled to fit all images. The idea is to measure said distribution once for each individual camera during production, compress it, and save it in a non-volatile memory. For each image that is then taken with the module the noise pattern can be retrieved and rescaled to fit the image

[0025] FIG. 1 describes the basic components of the system invented. Said system comprises a lens or a lens system 1, an image sensor 2, an optional image buffer 3, a processing device 4 using FPN data stored in a non-volatile memory 5 to eliminate the FPN of the sensor. The number 6 represents the corrected image. Said optional image buffer can be implemented by a RAM buffer.

[0026] FIG. 2 describes the basic steps of the method invented. In step 20 the FPN distribution of the FPN data are derived for each sensor using calibration frames. Said FPN data are stored in a non-volatile memory in step 21. Step 20 and 21 are performed only once for each sensor during production of the camera. In step 22 said FPN data are retrieved from said non-volatile memory for each picture taken and applied in step 23 on each pixel of the images taken to compensate Fixed Pattern Noise.

[0027] In a preferred embodiment quantization is employed to reduce the amount of data and a non-lossy compression scheme is used to minimize memory requirements. Quantization is the division of a quantity into a discrete number of small parts, often assumed to be integral multiples of a common quantity. The quantization level is defined in the quantization process as the discrete value assigned to a particular subrange of the signal being quantized. Depending on the quantization level used a reasonable compression results in the order of 2.5 bits per pixel. The black level is corrected for as part of the operation.

[0028] FIG. 3 describes the method used in a preferred embodiment of the invention to compensate Fixed Pattern Noise (FPN). In step 30 the FPN data are derived for each sensor using calibration frames. Said calibration frames are obtained at high temperatures with long exposure times with the camera lens covered. Said FPN data comprise the device and interconnect parameter variations (mismatches) across a sensor. In the next step 31 said FPN data are reduced by quantizing said parameter variations of the FPN data. Said FPN data are categorized into discrete number of categories. Step 32 and 33 deal with a compression of said quantized FPN data. In our preferred embodiment the Huffman compression algorithm has been used. Any other non-lossy compression method could be employed as well.

[0029] The Huffman compression algorithm is an algorithm to compress data. It does this by assigning smaller codes to frequently used symbols and longer codes for symbols that are less frequently used. A Huffman code can be represented as a binary tree whose leaves are the symbols that are encoded. At each non-leaf node of the tree there is a set containing all the symbols in the leaves that lie below the node. In addition, each leaf is assigned a weight (which is the frequency of the symbols), and each non-leaf node contains a weight that is the sum of all weights of the leaves lying below it. In step 32 an optimized Huffman table is generated, in other words a tree is defined, which will encode the FPN data with the fewest bits, and to save the symbols and the codes for each symbol along with the length of each code in a file in the form of a table. In step 33 the quantized FPN data are compressed using said optimized Huffman table. In step 34 said quantization factor used in step 31, the compressed FPN data, the Huffman table generated in step 32 and used in the compression step 33, and the distribution measurement is saved in a non-volatile memory. In a preferred embodiment the standard deviation a has been employed to describe said distribution. Other distribution parameters could be used as well.

[0030] Steps 30 to 34 are calibration steps. Said calibration is performed once for each sensor during production of said sensor. The following steps are application steps, performed for each image taken.

[0031] Step 35 describes the Huffman decompression of the FPN data stored in the memory. First, the Huffman tree is regenerated from the table stored in the memory as well. Once this process is completed a complete Huffman tree is obtained. Said tree is used to search for the character that was represented by the code in the compressed file. Using the regenerated Huffman tree the decompression is performed and the decompressed FPN data are directly used or saved in an output file.

[0032] In step 36 a quantization scaling is performed using the quantization factor stored in the non-volatile memory and the current statistical distribution measurement. The impact of temperature and exposure time on the fixed pattern noise is considered by a normalization of the stored data by statistical data obtained from “covered pixels”. “Covered pixels” are pixels having covered with metal or other opaque materials. In a preferred embodiment the data is normalized by the standard deviation σ of the calibration image and scaled by the standard deviation σ′ of the target image. Another statistical distribution measurement could be used as well.

[0033] Finally, in step 37 the scaled quantization data of step 36 are applied to each of the pixels of the image taken. The standard deviation σ of the covered pixels in the target image is used for the above-mentioned scaling and the average is used for blacklevel subtraction. The fixed pattern noise is compensated without the need of an additional external shutter and with low memory requirements.