WO2012160264A1

WO2012160264A1 - Method for carrying out a correction in the non-uniformity of a response from light-sensitive detectors

Info

Publication number: WO2012160264A1
Application number: PCT/FR2011/000312
Authority: WO
Inventors: Eric Prang
Original assignee: Eric Prang
Priority date: 2011-05-24
Filing date: 2011-05-24
Publication date: 2012-11-29

Abstract

Heat-detecting arrays provide images even without light, for example at night. Said detection arrays can be from a plurality of technologies (cooled or uncooled) and can be sensitive in a plurality of wavelengths. But in every instance, the detection arrays have disparities that render the images from heat cameras using said sensors non-uniform. Certain pixels, different from the ones adjacent to them, are lighter or darker on the image. Said non-uniformity, changing over time, necessitates the real-time correction thereof, if possible, without masking the image via a self-contained source. Said correction is generally called NUC (Non-uniformity correction). By means of a method using simple calculations, for example, integrated into a cabled logic circuit (FPGA) near the detector, the invention makes it possible to permanently correct the response of said detectors from a level lower than the intrinsic noise of the detector, even over a textured or untextured, moving or stationary, scene background. Said method depends on slow pixel-by-pixel correction table convergence updated by detection of extremely local pixels. Thus the method has no threshold or settings and operates for all detectors. Said method is particularly useful for implementing detection of very small targets in the images. The method is also of use after the fact on recorded image sequences, something impossible with conventional adjustment methods that use a self-contained reference source and oblige the user to adjust the camera before taking images.

Description

Title: A method for effecting non-uniformity correction of photosensitive detectors.

Summary :

The proposed invention makes it possible to make a correction of the non-uniformity of the matrix detectors of the thermal cameras continuously and on a landscape background, that the camera is in motion (on a carrier, in rotation for a panoramic watch, ...) or not. It is applicable to all detectors, cooled or not. A simple electronic, wired logic type, performs the detection of response errors of all the pixels of the image by comparison with the value of neighboring pixels. Local extremum detection is performed. A statistic principle, based on the probability of having a pixel, a level, strictly greater or smaller than its neighbors, as a function of the camera's spatial bandwidth (MRTD: Minimum Resolvable Temperature Difference) makes it possible to correct a fraction of the (offset) error detected on this pixel. The correction value, stored in a memory, is then adjusted. The statistical phenomenon being set up during the flow of images, the homogeneity defects of the various pixels are then progressively corrected. A detection of the dynamic range of the corrected pixel (hot point or cold point) and a storage of the previous corrections also makes it possible to correct the gain of the pixel. This method also makes it possible to detect the "dead" pixels of the detector and to make a "conventional" correction by replacing it with a neighboring pixel operating correctly. This method does not require adjustment and operates without threshold which makes it very robust.

Technical area :

The invention relates to the basic treatments of thermal images. These treatments are generally carried out by the manufacturers of thermal cameras, but can also be made by detection systems, automatic or not, or by visualization systems or presentation of images (pre or post treatments).

All technologies of thermal detectors is concerned, cooled or not. Even if the homogeneity errors are different according to the detectors, the process, being self-adaptive, can be used effectively.

State of the art :

At present the correction of non-uniformity of the (thermal) sensors is generally carried out by measuring a uniform source placed in front of the detector (shutter or source of reference, or defocusing of the image by the lens).

The method consists of exposing the sensor to as homogeneous a source as possible, measuring the result and storing the deviations in order to then provide an inverse correction. 1 000312

Some methods also offer corrections from video stream analysis on images made with landscape backgrounds (without reference source). It is then necessary to make an analysis of the evolution of an intensity radiated by a portion of the scene in case of displacement in the image. These methods, which are much more complex than the device that is the subject of the invention, generally presuppose restrictions on the nature of the observed scene.

(stationarity in time, ...).

The response of the detector pixels can be modeled approximately by a straight line: R ^ G _y xP ^ + O ^

Or G _y is the gain of the pixel located at the coordinates i and j in the detector matrix, Oy being the offset of this same pixel. The intrinsic error of the detector is then for the pixel ij the difference Rij-Pij

Exposure to a single temperature level does not allow the correction due to the Gy term, but on the other hand, this term Gy is generally much more stable in time, and therefore requires only a small correction variation.

Gy gain correction is achieved by making two exposures of the detector at two different temperatures. A linear regression type calculation then makes it possible to estimate all of the two parameters.

The variation over time of the parameter Oy depends on the technology of the detector used, but in all cases forces to make this correction relatively often (at each start, and approximately every hour according to the thermal variations of

the detector environment)

This operation disturbs the image pickup that is interrupted during the measurement of the responses of the elementary detectors.

In addition, the electronic and thermal noise disturb the pixel response measurement performed on the homogeneous source. These measurements are therefore performed on several images in order to average the measurements and to minimize these noises. This process is time consuming.

Finally, the "homogeneous source" is not perfect and generates with this process correction errors (at low spatial frequency) which can be detrimental to the quality of the image.

The sources (or references) external to the camera, see the internal shutter systems, are bulky and expensive. The invention that eliminates this need therefore allows space and cost savings on the camera.

Presentation of the invention

The invention relates to a device for real-time correction of inhomogeneity of thermal images, characterized in that it comprises:

• A matrix detector (of any technology, cooled or not,) placed behind an adapted optics,

• A digitizing system that provides a digital image with two spatial dimensions and a value representative of the detected temperature • Wired electronics (or possibly a conventional computer programmed) which includes a device for comparing values of neighboring pixels, calculation operators, parameter memories associated with each pixel.

• Possibly a non-volatile memory to memorize the corrections in case of power failure.

The invention also relates to an image analysis method which comprises the following steps:

1. the correction calculation of the image according to the calculated and stored correction tables

2. the measurement of local extremums in the image (strict comparison with close neighbors)

3. the modification of a step of the offset value of the pixel extremum (detected by step 2) in the opposite direction to its deviation from the neighbors

4. memorizing the value of the pixel, as well as the correction performed

5. the analysis of the different corrections made in the past and the error search on the pixel gain

6. the possible modification (according to step 5) of a step of the gain value of the pixel in the identified direction. Then the reset possible memories of corrections made on this pixel.

7. analyzing the pixel correction frequency, or offset error levels, and characterizing the pixel as "dead pixel". Then the correction of this pixel.

(See Figure 3: Flowchart)

This process is performed in real time on the images. The correction being progressive, the optimum is obtained only after a certain time, depending on the scene observed.

It can be observed that exposure to a homogeneous scene (classical method) makes it possible to accelerate the convergence of the correction, but is no longer necessary, thanks to this invention.

When the camera is turned off, the Offset and Pixel Gain correction tables can be memorized, so that they can be reused instantly at the next start, thus increasing the speed of convergence and thus the production of quality corrected images.

The correction made on the Pixel Offset is one step, which is the smallest unit measurable by the system. This level, lower than the electronic noise and the thermal noise of the images makes it possible to make a correction which thus leads to a correction better than these noises.

In the case of a homogeneous source, if one considers the correction stabilized in time, it is the noises which, added to the signal, will determine the pixels "local extremum". A pixel seen as too bright by the device will have its offset decreased by one step.

Statistically, if this fix was abusive, the pixel will appear in a time

indeterminate as too dark, and will benefit from a reverse correction. The correction error will therefore be at most one step, therefore less than the noise. The invention thus amounts to averaging the corrections in time, whereas the traditional method previously averages the images to eliminate the noise, and makes a correction thereafter.

If the source is not homogeneous, the operation of the invention is identical, except for sources that appear punctually in the image. If the camera and the point source (target) are perfectly fixed in time, the Offset of the pixel considered may be subject to improper correction to the nearest neighborhood pixels. This case is very rare for the following reason: the resolution of modern detectors is generally more important than the stabilization of the line of sight and the point source is not always displayed in the same place. Moreover this risk of degradation of the final image is not very important because in this case we are at the limit of resolution of the image. In the case of observation in a direction stabilized by an external means (gyrostabilization for example), a variant of the invention makes it possible to minimize the risk of occurrence of this defect at the cost of an increase in the filter convergence time. It is a question of interposing a step 2 bis to the process which consists in verifying that the pixel is extremum (of the same direction) in two successive images.

(See Figure 1: Pixel "statistically too bright")

The detection of an incorrect gain is made if a pixel on a dark source (cold object) is detected too dark and if the same pixel on a clear object (hot source) is detected too light or vice versa. It is therefore the natural variation of the visualized landscape that allows this very slow correction. This slowness is legitimate because the variation of the intrinsic gain of a pixel is small and very slow in time. The processing device therefore characterizes the pixels as dark or light according to an arbitrary and insignificant threshold. Pixels greater than 2/3 of the dynamics of the image will be considered as clear, pixels less than 1/3 of the dynamics will be considered dark. The other pixels

(included in 1/3 and 2/3) will not be considered. Each pixel has a clear correction memory (CC) and a dark correction memory (CS). These memories are initialized at startup with zeros. If an offset correction is detected for a pixel, the memory corresponding to its level is initialized to +1 if the Offset is increased, and -.1 if it is decreased.

If a CS = +1 and CC = -1 configuration is detected, the Pixel Gain is decreased by one step and the CS and CC memories are reset to 0.

If a CS = -1 and CC = +1 configuration is detected, the pixel gain is increased by one step and the CS and CC memories are reset to 0.

(See diagram 2: Correction of the offset then correction of the gain)

The dead pixels, that is to say having a constant response in time, will be detected by the device as local extremums at a much higher frequency than the others. The device therefore comprises a conventional counter system (incremented by the corrections and decremented regularly) or another conventional frequency measurement system in electronics or signal processing to detect these pixels.

A specific memory will then be switched to a state storing this detection. The value of the pixel will then be replaced by the value of the last processed pixel having no defect. The advantages of this invention are that it allows detection of nonuniformity errors without adjustment (without threshold) and without tooling or external means (homogeneous sources). Furthermore the correction is performed in real time and without disturbing the image flow that can be used continuously for viewing or for other treatments (tracking targets, automatic detection, ...).

This method can also be used a posteriori on a recorded image sequence. A first pass on the images (or part of the images) making it possible to generate a table of correction of offsets and possibly gains of each pixel. This table is then used to correct the sequence or the entire film. This method therefore makes it possible to correct images a posteriori, even if an evaluation of the nonuniformity defects of the sensor has not been carried out before the image is taken. These corrections a posteriori not being feasible in the case of correction of conventional Uniformity by visualization of a homogeneous reference scene.

Claims

6 Claims:

1. Method for real-time non-uniformity correction of matrix detectors

used to make thermal images characterized by the following steps:

at. Correction of the nonuniformity of each pixel by summation with an offset correction table.

b. Detection of local extremum pixels by comparison of their value with respect to the value of neighboring neighboring pixels,

vs. Change the value of the correction table for extremum pixels by one step in the opposite direction of the deviation from the neighbors.

2. Method according to claim 1, which also corrects the gain defects of each pixel and characterized by the addition of the following steps:

at. Memorization of corrections made in time to correct the gain assigned to each pixel in two tables: one for the light pixels and one for the dark pixels

b. Detection of opposite corrections in the two dark and clear tables and in this case modification of the gain by increase or decrease of a very low value, then reset of the clear and dark tables.

3. Method according to claim 1 which detects the "dead" pixels (insensitive to

landscape) and characterized by adding a step of detecting a significant (and abnormal) consecutive correction frequency of a pixel.

4. The method of claim 1 which accelerates the convergence of the values of offsets to the values of the fixed structure of the sensor without removing any details of the scene in a still image (stationary camera) by modifying step c of claim 1 and characterized by the correction value of the pixel offset which is different as a function of time (more important for the first images) or as a function of the background scan, detected by movement in the image or by a sensor external.

5. The method of claim 1 which performs a fast correction from the first images and characterized by the addition of a step of initializing the offset correction values to stored values during previous correction.

6. The method of claim 1 but out of real time and characterized in that it performs the correction a posteriori on an image sequence. The table of offset correction values and possibly gain values according to claim 2 is then applied to all the images.

7. Device which implements the method cited in claims 1 to 5 by

the use of electronic circuits based on simple functions (comparators, memories, adders, etc.) or by the use of an FPGA type integrated circuit.

8. Device according to claim 6, and characterized by the use of a nonvolatile electronic memory of the size of the pixel number in the images to keep in time and even after power off the camera, the data. correction. At startup, after powering up the corrections

(offset) are initialized by this memory.

9. A computer program that performs a nonuniformity correction by performing the steps of the method mentioned in claims 1 to 6