IES84135Y1 - Method and apparatus for red-eye detection in an acquired digital image - Google Patents

Description

Method and Apparatus for Red—Eye Detection in an acquired Digital Image FIELD OF THE INVENTION The present invention relates to digital image processing. and more particularly to a method and apparatus for red—eye detection in an acquired digital image.

BACKGROUND TO THE INVENTION Red-eye is a phenomenon in flash photography where a flash is reflected within a subject's eye and appears in a photograph as a red dot where the black pupil ofthe subject's eye would normally appear. The unnatural glowing red of an eye is due to internal rellections from the vascular membrane behind the retina. which is rich in blood Vessels. This objectionable phenomenon is well understood to be caused in part by a small angle between the llash olthe camera and the lens ofthe camera. This angle has decreased with the miniaturi/ation o I‘ cameras with integral flash capabilities. Additional contributors include the relative closeness otthe subject to the camera and ambient light levels.

Digital cameras are becoming more popular and smaller in size. US 6.407.777 to I)eI_uca describes a method and apparatus where a red eye filter is digitally implemented in the capture device. The success or failure of such filter relies on the quality ofthe detection and correction process.

Most algorithms that involve image analysis and classification. are statistical in nature There is therefore a need to develop tools which will improve the probability ofsuccessful detection. xx hile reducing the probability of false detection. while maintaining optimal execution. especially in limited computational devices such as in digital cameras. In many cases knowledge of the image characteristics such as image quality may affect the design parameters and decisions the detection and correction software needs to implement. I-‘or example an image with suboptimal exposure may deteriorate the overall detection of red-eye detects.

Thus. what is needed is a method of improving the success rate of algorithms tor detecting and reducing red-eye phenomenon.

S84l 35 SUMMARY OF THE INVENTION According to the present invention there is provided a method and apparatus for red-e}e detection in an acquired digital image as claimed in the appended claims.

The present invention compensates for sub—optimally acquired images where degradations in the acquired image may affect the correct operation of redeye detection. prior to or in conjunction with applying the detection and correction stage.

The present invention improves the overall success rate and reduces the false positive rate of red eye detection and reduction by compensating for non-optimally acquired images by performing image analysis on the acquired image and determining and applying corrective image processing based on said image analysis prior to or in conjunction with applying one or many redeye detection filters to the acquired image. Such corrections or enhancements may include applying global or local color space conversion. exposure compensation. noise reduction. sharpening. blurring or tone reproduction transformations. in preferred embodiments. image analysis is performed on a sub—sampled copy of the main acquired image where possible. enhancing the performance ofthis invention inside devices with limited computational capability such as hand held devices and in particular digital cameras or printers.

In the preferred embodiment. the pre—tiltering process is optimized by applying when possible, as determined from the image analysis, the image transformations at the pixel level during the redeye detection process thus compensating for non—optimall_v acquired images without requiring that corrective image processing be applied to the full resolution image.

In preferred embodiments. the redeye filter chain is configured for optimal performance based on image analysis of an acquired image to enhance the execution red eye detection and reduction process. Such configuration takes place in the form of variable parameters for the algorithm and variable ordering and selection ofsub—lilters in the process.

Preferred embodiments of the invention operate uniformly on both pixels which are members of a defect and its bounding region thus avoiding the need to determine individually if pixels in the neighborhood of said defect are members of the defect and to subsequentl_v apply correcting algorithms to such pixels on an individual basis.

Using preferred embodiments of the present invention, variables that could significantly effect the success of the red—eye detection algorithm such as noise. color shifts. incorrect exposure. blur, over sharpening etc. may be pre-eliminated before performing the detection process. thus improving the sticcess rate.

Alternatively or in addition these yariables may be pre-accounted for by changing the parameters for the detection process, thus improving the performance and the success rate.

An advantage of preferred embodiments of the present invention is that by bringing images into a known and better defined image quality. the criteria for detection can be tightened and narrowed down, thus providing higher accuracy both in the positive detection and reduction in the false detection.

A further advantage of preferred embodiments of the present invention is that by accounting for the reasons for suboptimal image quality the parameters for the detection and correction algorithm may be modified. thus providing higher accuracy both in the positive detection and reduction in the false detection without the need to modify the image.

An additional advantage of preferred embodiments of this invention is that misclassification of pixels and regions belonging to defect areas is reduced if not altogether avoided. which means a reduction of undetected correct positives.

An additional advantage of preferred embodiments of this invention is that color misclassification of pixels and regions belonging to non—defect areas is reduced if not avoided. which means a reduction of false positives.

A further advantage of preferred embodiments ofthe present invention is that they can be implemented to run sufficiently fast and accurately to allow individual images in a batch to be analyzed and corrected in real-time prior to printing.

Yet a further advantage of preferred embodiments of the present invention is that they have a sufficiently low requirement for computing power and memory resources to allow it to be implemented inside digital cameras as part ofthe post-acquisition processing step.

Yet a further advantage of preferred embodiments of the present invention is that they have a sufficiently low requirement for computing power and memory resources to allow them to be implemented as a computer program on a hand-held personal digital assistant (PDA). mobile phone or other digital appliance suitable for picture display A further advantage of preferred embodiments of the present invention is that they are not limited in their detection of red-eye defects by requirements for clearly delinetl slxin regions matching a human face.

A further advantage of this invention is the ability to concatenate image quality transformations and red eye detection to improve overall performance.

BRIEF l)l{SCRlP'FlON OF THF. DRAWll\l(iS Fig. 1(a) shows a prior art in—camera redeye detection system: Fi . l(b) shows an improved redeye detection system according to an embodiment of the present invention; Fig. 2(a) is a flowchart illustrating the operation ofthe system of Figure l(b); Fig. 2(b) is a flowchart illustrating an alternative mode of operation ol‘ the system of Figure 1(b); Fig. 2(c) illustrates another alternative mode ofoperation of the system of Figure l(b): Fig. 2(d) is a flowchart illustrating a further alternative mode of operation of the system ofFigure l(b); Fig. 2(e) is a flowchart illustrating a still further alternative mode of operation of the system of Figure l(b); Fig. 3 shows the redeye filter chain of Figure l(b) in more detail: Fig. 4(a) illustrates the operation of portions of Figures 2(a). 2(b). Zttl) & 2(e) in more detail: 4(b) illustrates an alternative implementation ofFigure 4(a): Fig. 4(c) is a flowchart illustrating the operation ofa portion of the system of Figure l(b); and Figs 5(a) and 5(b) illustrate the operation of a red-eye filter chain according to an embodiment of the present invention.

[U U1 DETAILED DESCRIPTION OF THE lN\/HNTION I’ig 1 illustrates a prior art in—camera redeye system. Within the camera 100 a main image is acquired 105 from a sensor subsystem. This image is further processed 110 based on image acquisition parameters such as ambient lighting, length of exposure. usage of pre-llash and flash. lens focal length & aperture settings. etc. This image processing is pre-calibrated during the design of the camera and, due to the non-linear relationships between the various acquisition parameters. it typically involves a significant amount of empirical testing using as broad a range of image capture conditions as is practical. Thus, even though modern digital cameras lia\'e much improved auto—focus and auto—exposure algorithms it is still possible to capture images of non- optimal quality either through incorrect camera settings or through encountering conditions which are not fully accounted for by the empirical calibrations process for that camera.

After this image processing is completed the main acquired and processed image is normally committed to non-volatile storage in camera memory. or in an oiiboard storage card 170. However if the image was captured using a flash then the possibility of rede_\e defects implies that the image should first be passed through an in—camera redeye filter 90. A more detailed description of such a filter can be found in US 6,407,777 to DeLuca herein incorporated by reference. Briefly it comprises of (i) a pixel locator filter 92 which detects candidate eye- defect pixels based on a color analysis and then groups said pixels into redeye candidate regions: (ii) a shape analyzer filter 94 which determines ifa eye candidate region is acceptable in terms of geometry. size and compactness and further analyzes neighbouring features such as eyebrows and iris regions; and (iii) a falsing filter 98 which eliminates candidate regions based on it w ide range of criteria. Any candidate regions which survive the falsing filter are then modified by a pixel modifier 96 and the corrected image 170-2 may then be stored in the main image store 170.

This prior art system typically will also feature a sub-sampler which can generate lower resolution versions 170-3 of the main acquired and processed image 170-1. This sub—sampling unit may be implemented in either software or may be hardware based and is. priiiiarily. incorporated in modern digital cameras to facilitate the generation of thumbnail images for the main camera display.

[Q ‘J1 Fig l(b) illustrates a preferred embodiment of red—eye detection system according to the present invention. The system improves on the prior art by providing an additional image cilia/y.s‘i.s‘ prefi/(er 130 and an image Conzpemuiion prtffi/fer 135 to the prior art imaging chain to reduce the overall incidence of errors in the redeye detection process 90 for non—optimall_y acquired images.

The image analysis prefilter 130 combines one or more techniques for determining image quality. Such techniques are well known to one familiar in the art of image processing and in particular image editing and enhancements. Thus, the prefilter provides an in—camera analysis of a number of characteristics of an acquired. processed image with a view to determining it‘ these characteristics lie within acceptable limits. It will be clear to those skilled in the art that the exact combination of analysis techniques will be dependent on the characteristics of the non—optimally acquired images generated by a particular digital camera. In addition. the determination ol‘ what image quality matters need to be addressed is primarily dependent on the etiect of such characteristics on the red eye filter 90. Thus. as illustrative examples: (i) a low-end digital camera may omit complex noise filtering circuitry on its sensor as it is targeted at cost—sensitive markets and may employ low quality optics tor similar reasons. Thus it may be susceptible to a greater degree of image noise and exhibit a poor dynamic range for white and color balance; (ii)a high—end professional camera will have a much greater dynamic range lor color and vrhite balance but may require more sophisticated image analysis to compensate for motion blur. sensor dust and other image distortions that are ot‘ concern to professional photographers.

Accordingly we shall provide some examples of image analysis techniques for exemplary purposes only and it will be understood these are not intended to limit the techniques which may be utilized in implementing the present invention. ()ne subsystem of the image ana/y.s'i.s' prefiller is a blur analyzer 130-], which pcrlorms an image analysis to determine blurred regions within a digital image — this operate on either the full size main image 170-1 or one or more sub-sampled copies oi‘ the image 170-3. One technique for in—camera blur detection is outlined in US patent application 2004/0120598 to Feng which describes a computationally efficient means to determine blur by analysis of DCT coellicicnts in a JPEG image. In common with the other sub—s_vstcms of the prefilter 130. the analyser provides a measure of the blur in the supplied imagets) to be used later in the prefilter 135. lhis measure could be as simple as an index between 0 and 1 indicating the degree of blur. However. it could also indicate which regions of the image are blurred and the extent to which these are blurred.

A further subsystem of the image unu/y.s'i.s' prefi/Ier is a dust analyzer I30-2. lihe problems caused by dust on imaging devices are well ltnown in the prior art. In the context of the present invention it is important to track the location and severity of dust particles as these may interfere with the correct detection of eye—defects when the two forms of defect oxerlap. Of particular relevance are techniques where the detection of defects in a digital image is based solely on analysis of the digital image and that do not directly relate to the image acquisition process. For example US patent 6,233,364 to Krainiouk et al. discloses determining anomalous image regions based on the difference between the gradient of an image at a set of grid points and the local mean of the image gradient. This technique generates few false positives in ‘‘noisy'' regions of an image such as those representing leaves in a tree. or pebbles on a beach. US patent 6,125,213 to Morimoto discloses detecting potential defect or "trash" regions within an image based on a comparison of the quadratic differential value of a pixel with a pre-determined threshold value. In addition. Morimoto discloses correcting "trash" regions within an image by successively interpolating from the outside of the "trash" region to the inside of this region ~ although this does not need to be performed by the subsystem 130-2. US patent 6.266.054 to Lawton et al. discloses automating the removal of narrow elongated distortions from a digital image utili7.ing the characteristics of image regions bordering the distortion. US patent application 2003/0039402 and WIPO patent application W0-03/01 94 73 both to Robins er al. disclose detecting defective pixels by applying a median filter to an image and subtracting the result from the original image to obtain a difference image. This is used to construct at least one defect map and as such provide a measure of the effect of dust on an image supplied to the subsystem I30-2.

US patent 6,035,072 to Read discloses mapping defects or dirt. which affect an image acquisition device. A plurality of images are processed and stationary components \\hich are common between images are detected and assigned a high probability of being a defect.

Additional techniques which are employed to modify defect probability include median filtering. _.

(J1 sample area detection and dynamic adjustment of scores. This dynamic defect detection process allows detect compensation, defect correction and alerting an operator ol‘ the likelihood ol‘ defects. but from the point of View of the preferred embodiment, it is the map which is produced which indicates to the prelilter 135 the degree to which the supplied images are affectetl by dust and/or defects.

Additional subsystems of the image (Incl/_V.s‘i.\‘ prefi/{er are a white balance analy /er 130-3. a color balance analyzer 130-4, and a gamma/luminance analyzer 130-5. In the embodiment. each of these provides. for example, an indicator of the degree to which each of these characteristics deviates from optimal and by which the supplied image might be corrected. Those skilled in the art will realize that such techniques are practiced in a digital camera as part ol‘ correeti\c image processing based on acquisition settings 110. Prior art techniques which can be employed in embodiments of the present invention also exist for post-processing of an acquired image to enhance its appearance. Some representative examples are now described: US patent 6,249,315 to Holm teaches how a spatially blurred and sub—sampled version ol‘ an original image can be used to obtain statistical characteristics of a scene or original image. In Holm. this information is combined with the tone reproduction curves and other characteristics of an output device or media to provide an enhancement strategy for digital images. whereas in the preferred embodiment. an analysis prefilter employing the technique ol‘ Holm prelerabl_\ provides the color characteristics of the supplied image to the prefilter 135.

US patent 6,268,939 to Klassen et al. teaches correcting luminance and ehrominancc data in digital color images. Specifically. Klassen is concerned with optimi7ing the l1‘€t11Sll)l‘l11fi1l0l1S between device dependent and device independent color spaces by applying subsampling of the luminance and ehrominancc data.

US patent 6,192,149 to Eschback et al. discloses improving the quality ol‘ a printed image by automatically determining the image gamma and then adjusting the gamma ol‘ a printer to correspond to that of the image. Although lischback is concerned with enhancing the printed quality of a digital image and not the digital image itself. ifdoes teach a means for au1omaticall_\ determining the gamma of a digital image and as such can be used in an analysis pre-lilter in embodiments of the present invention. US patent 6,101,271 to Yamashita et al. discloses implementing a gradation correction to an RGB image signal which allows image brightness to i\) ‘J1 be adjusted without affecting the image hue and saturation A further subsystem of the image (Ina/ysis prqfi/lei‘ is an image texture analy/.er 130-6 which allows texture information to be gathered from the acquired and processed main image.

This information can be useful both in determining different regions within an image and. when combined with information derived from other image analysis filters such as the blur analyzer 130-1 or the noise analyzer 130-7 it can enable automatically enhancement of an image by applying deblurring or denoising techniques. US patent application 2002/0051571 to Jackway et al discloses texture analysis for digital images. US patent application 2002/0090133 to Kim et al discloses measuring color—texture distances within a digital images and thus offering improved segmentation for regions within digital images.

A further subsystem of the image aria/ysis‘ prcffi//er is a noise analyzer 130-7 The last illustrative subsystem of the iinagc (ma/_i:s'i.s' prcjfi//er 130 is an objeet’region analyzer 130-8 which allows localized analysis of image regions. One particular region which will invariably be found in an image with eye-defects is a human face region. lhe detection ofu face region in an image with eye—defects is simplified as described in [IS patent application 2004/0119851 to Kaku. Again, an analysis pre—lilter employing Kaku would there provide indicators of where faces regions are to be found in a supplied image to the pre-filter I35.

Turning now to the image con1;2ensali(iii prtffi/lei’ 135. In the present embodiment. a combination of image correction analyzer 135-2 and a rcdcire .921/ifi/lei" dala/msc 135-3 (i) interpret the results of the image analysis performed by the iiiicigc (Illa/_l‘.\‘l'.\ /ire/i/Icr I 30; (ii) if corrective image processing is active. determine an optimal correction strategy for application to the acquired. processed image. or a subsainpled copy thereof; (iii) if adaption of the redeye filter chain is implemented. determine any parametcr’f1lter conflicts and further determines an optimal adaption of the redeye filter chain (described later): and (iv) if both corrective image processing and filter adaption are active. determine an optimal combination of each.

‘J I Lu © The actual crirrecfive image pr0ce.s;s'z'ng 135-1 will typically be implemented as a library of image processing algorithms which may be applied in a variety of sequences and combinations to be determined by the image correcliarz anal_t>zer 135-2. In many digital cameras some olithese algorithms will have partial or full hardware support thus improving the peitormance oi‘ the c*ompen.s‘aI/on prefiller 135.

It was already remarked that the ana/y.s‘is p/‘e_filter 130 can operate on a subsampled copy of the main image 170-3. In the same way the detection phase of the redeye tilter 90 can be applied to a subsampled copy of the main image 170-3. although not necessarily ol‘ the same resolution. Thus where corrective image processing is used by the image co/i1pcmaI1'mz prefi/Iai- it will also be applied to a subsamplcd copy of the main image 170-3. This has significant benetits with respect to computation speed and computing resources. malxing it particularly advantageous for in—camera embodiments.

We also remark that the image carrectimi analyzer 135-2 may not always be able to determine an optimal correction strategy for an acquired. processed image due to conllicts between image processing algorithms, or between the filter adaptions required tor the redeye tiller chain. In other instances, where a strategy can be determined but the image (‘U/‘/‘L’L‘Il()I7 analyzer 135-2 may be aware that the strategy is marginal and may not improve image quality it may be desirable to obtain user input. Thus the image correction analyzer 135-2 may generate a ziser imlicalirm 140 and in certain embodiments may also employ additional user interaction to assist in the image correction and redeye filter processes.

Fio, 2a to Fig 2e illustrate several alternative embodiments of the present im emion which are described as follows: (i) In Fig 2(a) an acquired. processed main image. or alternatively a subsampled copy thereof, is initially loaded, step 201 to respective sub—systems of the anal}‘.s'i.\' prcfil/er /3/J, slap 202. These produce their measurements and a determination is made ilany ol‘ the image quality characteristics lie within or outside acceptable thresholds is made by the image correction analyser 135-2. step 204. If image quality is within acceptable limits lot‘ each of the image characteristics analyzed then the redeye filter 90 can be applied normally and no corrective image processing is required. However. if certain image characteristics do lie outside acceptable tolerances then additional analysis is performed by the analyser I35-2 to determine if corrective image processing can be applied 206. If some ofthe analy/ed image characteristics lie too far outside acceptable thresholds. or if a disadvantageous combination of image characteristics is determined. it may not be possible to correct the image reliably prior to applying the redeye filter. Thus the filter 90 can be disabled 220. a user indication can be provided and processing is completed for this particular image 224. xyithout performing the red eye correction or performing the process with lower probability of success. However. if the image can be repaired. 206-YES. the image is corrected step 208. prior to executing the red eye algorithm 90. In the preferred embodiment. the process of correcting the image. 208 may be performed on the full resolution image. or alternatively on a subsampled image or a copy of the image. The exact nature and possibilities for such corrections. 208. whether locally or globally are described later. In any case. the corrected image needs only be stored temporarily and can be discarded after red-eye processing is complete. 209. It should be noted that performing the pre-filtering. 208 on the image. does not means that the actual red—eye detection and reduction algorithm. 90 has to be modified to account for possible variability. Nonetheless. as image quality supplied to the lilter ‘)0 is improved, the red eye algorithm can use tighter parameters and more well defined restrictions as to the nature of the red eye features that are to be identified so producing improved results. (ii) Fig 2(b) corresponds with figure 2(a) except that it includes an additional determining step. 240 which follows the determination that corrective image processing is possible. 206.

This additional step determines if the corrective image processing to be applied to the image can be provided by a globally applied transformation ofthe image pixels. The most popular global transformations are matrix multiplication or Iookup table transformations. Iiior example, the analysis provided by lilters I30-3 . . . l 30-5 may indicate to the analyser l35—2 that the principle cause of image non—optimality is a reddish color cast. In this case. a simple transformation of the red image component. R -—> R’ is sufficient to compensate for the image non—opti1nality. Another example will be an image that is under exposed and a tone reproduction curve (TRC) needs to be corrected. Global transformations have the advantage of being relatively computationally efficient and with a potential to be highly optimi7.cd. ln addition, such transformations may be performed within the redeye filter 90 itself. for example, as part of the pixel locator and region segmentation process 92 described in more detail later in relation to Figures 3 and 5. so reducing the overhead involved in performing this correction. For the moment, it is sufficient to say that in step 242. a pixel transformation within the pixel locator and region segmentor 92 of the red—eye filter is conligured. [t \x ill also been seen that the steps 240. 242 may be performed as an alternative to other corrections step 208. in parallel with other corrections or in series with other corrections prior to execution ofthe red-eye filter 90. (iii) In Fig 2(c) instead of corrective image processing to compensate for a non—optimally acquired image. the analyser 135-2 adapts the redeye filter chain to provide image compensation for the redeye detection process. Steps 202 and 204 are identical with previous embodiments. However. when an image characteristic lies outside acceptable tolerances. the analyser 135-2 determines the redeye sublilters affected by said out—of— tolerance image characteristic. step 250. Typically this determining step will involve the image correction analyzer 135-2 obtaining the relevant data from an in—camera data repository such as the redeye su/Jfi/z‘er data/7c/.ve I35-34 After the affected subtilters have been determined 250, the next step is to determine if subfilter compensation is possible 252.

This will depend on the different image characteristics which are outside acceptable thresholds and the relevant sets of redeye subfilters affected by each out—of-tolerance image characteristic. If filter chain adaption is possible then the filter chain is modified 254 and the redeye filter is applied 90. If subiilter compensation is not possible due to filter. or para1neter—based conflicts then steps 220, 140, and 224 are performed as in the previous embodiments. The subfilter determining process is further described in Fig 4(b) and an overview ofthe redeye subfilter matrix is given in Fig 3.

The following example illustrates the concept of the applying the results of the analysis stage to modify the filter chain of the correction process and the red eye detection process as opposed to modification of the image pixels. It is assumed that a pixel :I{(i.(ii..l3u! attcr the needed correction. step 208. is transformed to pixel value {Ri.Gt.Bi} by a ll‘itl]ST0l‘l]lLlllt)I1 '1‘: T[{R(,.G(,.B.,}] = {Rl.G,.B,}. For illustrative purposes. we assume that the tirst stage ol‘ the red eye detection algorithm, as defined in block 92 of Fig. 1(a) is comparison to a known value. to determine if the pixel is. in simplified terms. red or not. The Value of the pixel in to compare with is {R".G’.B’}. However. the two steps aboye of correcting and comparing may be combined simply by transforming the static Value of {R'.(i".B‘} based on the inverse of the correction transformation. Thus, thee preliminary preparator_\' stage will be: {R".(i‘i.B"} : '/"'[{Ri.Gi.Bi}j and the pixel by pixel comparison. as adapted. step 254 to the necessary needed transformations will comprise the following test: /FiRi.~Go~Btti 3 {RH-.G"-BU}. By doing so. the entire image is not corrected. but the comparison is similar to the state as if the image was corrected. The complexity and number of necessary steps compared to the original algorithm is exactly the same. with the extra value that the image algorithm now is taking into account the sub-optimal t]Lldll1_\' ol‘ the image. l\/lathematically speaking: T[{R(,.(j(,. B.,}]a{R',G',B‘} : {R0.o},.B., :aT"‘[{R',(;‘, Bat] : :R.),(;...B., :a:R'.<; '. /3‘: Where 0. denotes the relationship between the objects.

OI‘ course. such adaptation may be more complex than the simplilied example aboxve. and may include change of multiple values in the algorithm or change in the order the Various liltcrs are applied. or change in the weight of the Various filters. However. the improvement in performance may justify the added architectural complexity. (iv) Fig 2(d) illustrates a combination of the embodiments described in 2(b) and Ztc). This embodiment is identical to the previous embodiments except that if subtilter compensation is not possible 252 it incorporates two additional steps to detcrming it‘ correctixe image processing can be applied 206 and if this is possible a second step 208 to apply said corrective image processing. Note that subfilter adaption is preferred to corrective image processing as it requires practically no computational resources. but only changes the input parameters of the subfilters which comprise the redeye filter chain and the composition and ordcr—of—execution of the chain itself. However in certain circumstances correction of the original acquired image by image processing means may provide more reliable redeye detection, or be desirable as an end in itself. (v) Fig 2(e) describes an alternative variation of the algorithm. This is identical to the embodiment of Fig 2(a) except that after determining if corrective image processing is possible 206. corrective image processing is applied to both the main acquired image 170-] and a subsampled copy 170-3 thereof, step 208-]. A second additional step then saves the corrected acquired image 170-2. in the main image store 170, step 209, and a user indication 140 is generated to inform the camera user that an improved image is available. /\dditional steps may be added to allow the user to select between original l70—l and corrected images 170-2 if so desired. In this embodiment, redeye detection 92, 94, 98 is applied to the corrected subsampled copy of the main acquired image and the redeye correction 96 is applied to the corrected copy of the main acquired image. In other embodiments corrective image processing would not be applied to the full-sized main image 170-] so that the redeye correction would be applied to the zmcurrecled main image.

Fig 3 shows the principle sublilter categories which exist within the main redeye tilter 90.

While each of the component filters will be referred to in sequence. it will be appreciated that where appropriate more than one of these filters may be applied at a given time and the decisions above to modify the filter chain can include a decision not alone as to which lilters may be executed in a sequence, but also on which filters can be applied in parallel sequences. As described above. the pixel transformer filter 92-0 allows global pixcl—level transformations of images during color determining and pixel grouping operations. Also. within the pixel locator and region segmenter 92 we find pixel color filters 92-] which perform the initial determining if a pixel has a color indicative of a llash eye defect; a region segmcntor 92-2 which segments pixels into candidate redeye groupings; regional color filters 92-3. color correlation filters 92-4. and color distribution filters 92-5 which operate on candidate regions based these criteria. In addition the pixel locator and region segmenter 92 contains two additional functional blocks which do not contribute directly to the color determining and segmentation operations but are nevertheless intertwined with the operation of the pixel locator and region segmenter. The resegmcntation engine 92-6 is a functional block which is particularly useful for analyzing difficult eye defects. It allows the splitting 92-621 and regrouping 92-6h of borderline candidate regions based on a variety of threshold criteria.

After candidate eye—def‘ect groupings have been determined by the scgmenter 93. a shape analyzer 94 next applies a set of subfilters to determine is a particular candidate grouping is physically compatible with known eye—defects. Thus some basic geometric filters are first applied 94-1 followed by additional filters to determine region compactness 94-2 and boundary continuity 94-3. Further determining is then performed based on region size 94-4. and a series of‘ additional lilters then determine if neighbouring features exist which are indicative of eye shape 94-5. eyebrows 94-6 and iris regions 94-7. In certain embodiments of the present invention the redeyc filter may additionally use anthropometric data to assist in the accurate determining of‘ such features.

Now the remaining candidate regions are passed to a falsing analyzer 98 yyhich contains a range of subfiltcr groups which eliminate candidate regions based on a range ofcriteria including lips lilters 98-1. face region filters 98-2. skin texture filters 98-3. eye-glint filters 98-4. white region filters 98-5. region uniformity filters 98-6. skin color filters 98-7. and eye—region falsing lilters 98-8. Further to these standard lilters a number of specialized lilters may also be included as part ofthe falsing analyzer 98. In particular we mention a category of filter based on the use of acquired preview images 98-9 which can determine ifa region was red prior to applying a flash.

This particular filter may also be incorporated as part ofthe initial region determining process 92, as described in co-pending US application no. 10/9l9.226 from /\ugust. 2004 entitled "Red- Eye Filter Method And Apparatus" herein incorporated by reference. An additional category of falsing filter employs image metadata determined from the camera acquisition process 98-11).

This category offilter can be particularly advantageous when combined with anthropometric data as described in PCT Application No. PCT/EPZOO4/008706. Finally an additional category of filter is a user confirmation filter 98-11 which can be optionally used to request a final user input at the end of the detection process. This filter can be activated or disabled based on how sub- [J U1 optimal the quality of an acquired image is.

The pixel modifier 96 is essentially concerned with the correction of confirmed redeye regions and will not be discussed in greater detail here. In the preferred embodiment. an additional component of the redeye filter 90 is a filter chain adapter 99. This component is responsible for combining, and sequencing the subtilters ofthe redeye filter ‘)0 and for activating each filter with a set of input parameters corresponding to the parameter list(s‘) 9‘)-1 supplied from the image compensation prefilter 135.

Finally. it is remarked in the context of Fig 3 that although the pixel locator & region segmenter 92, the shape analyzer 94 and the liaising analyzer 98 are illustrated as separate components it is not intended to exclude the possibility that subtilters from these components may be applied in out—of—order sequences. As an illustrative example. regions which pass all the falsing lilters except for the region uniformity filter 98-6 may be returned to the resegmentation engine 92-6 to determine ifthe region was incorrectly segmented. Thus a sublilter from the pixel locator and region scgmentor 92 may be used to add an additional capability to the lalsing analysis 98.

Fig 4 shows in more detail the operation of the image analysis 130 and image compensation prefilters 135. In this example the operation of the compensation prelilter 135. and more particularly the operation of the image correction analyzer 135-2 has been separated into two functional modes: Fig 4(a) illustrates the workflow for the determining and perlmining corrective image processing (so corresponding generally to steps 206. 208 ol‘ Figure 2(a).(b),(d) and (e)) while Fig 4(b) describes the determining and performing filter chain adaption including determining if a single chain. or a combination of multiple lilter chains will compensate for the non-optimal image characteristics determined by the image analysis pretilter 130 (so corresponding generally to step 250,252 and 254 of Figures 2(c) and 2(d)). Fig 4(e) illustrates an exemplary embodiment of the workflow of the image analysis prelilter 130.

In Fig 4(a) the image correction analyzer 135-2 first loads an image characteristic list 40] obtained from the image analysis prefilter 130. This list will allow the correction analyzer to quickly determine il‘a simple image correction is required or ifa number of image characteristics will require correction 402. In the case of a single characteristic the correction analy7.er can immediately apply the relevant corrective image processing 412 followed by some tests oi‘ the corrected image 414 to ensure that image quality is at least not deteriorated by the applied corrective technique. If these tests are passed 416 then the image can be passed on to the redeye filter 90 for eye defect correction. Otherwise. if corrective image processing has failed the sanity tests 416 then an additional test may be made to determine iflilter chain adaption is possible 422.

In this case the algorithm will initiate the workflow described in Fig 4(b) for determining the required filter chain adaptions 450. If corrective image processing has failed 416 and filter chain adaption is not possible 422 then the correction analyzer will disable the redeye filter tor this image 220. and provide a user indication to that effect 140 after which it will pass control baclg to the main in—camera application 224. Note that in certain embodiments the user indication may be interactiye and may provide an option to allow the normal redeye filter process to proceed on the uncorrected image. or alternatively offer additional user—sele.ctable choices for additional image analysis and/or correction strategies. .\low returning to the determining step between single and multiple image characteristics requiring correction 402 we now describe the correction approach for multiple image characteristics. Typically an image which was non-optimally acquired will sutter from one major deficiency and a number of less significant deficiencies. We will refer to these as p/'imur_i' and .s'cc'0na’ar)' image deficiencies. The next step in the worktlow process is to determine the primary image deficiency 404. After this has been successfully determined from the image characteristics list the next step is to determine the interdependencies between this primary correction required and said secondary image characteristics. Typically there will be more than one approach to correcting the primary image characteristic and the correction analyzer must next determine the effects of these alternative correction techniques on the secondary image characteristics 406 before correction can be initiated. If any of the secondary characteristics are likely to deteriorate significantly and all alternative correction technique for the primary image characteristic are exhausted then the correction analyzer may determine that these interdependencies cannot be resolved 408. ln the present embodiment an additional test is next made to determine if filter chain adaption is possible 422. In this case the algorithm will initiate the workfiow described in Fig 4(b) for determining the required filter chain adaptions 450. If correctixe image processing has failed 416 and filter chain adaption is not possible 422 then the correction analy7.er will l\) 'Jl disable the redeye filter for this image 220, and provide a user indication to that ellect 140 after which it will pass control back to the main in-camera application 224.

Given that the secondary interdependencies can be resolved 408 the correction analyzer next proceeds to determine the image processing chain 410. In certain embodiments this step may incorporate the determining of additional corrective techniques which can il11‘1l1€I‘ enhance the primary correction technique which has been determined. In such an embodiment the correction analyzer will, essentially. loop back through steps 404, 406. and 408 for each additional correction technique until it has optimized the image processing chain. It is liirther remarked that the determining of step 408 will require access to a relatively cotnplex knowledgebase 135-4. In the present embodiment this is implemented as a series oi‘ look-up- tables (l,UTs) which may be embedded in the non—volatile memory of a digital camera. The content ol‘ the knowledgebase is highly dependent on (i) the image characteristics determined by the image analysis prefilter and (ii) the correction techniques available to the compensation prefilter and (iii) the camera within which the invention operates. Thus it will be evident to those skilled in the art that the knowledgebase will dilTer significantly from one embodiment to another. It is also desirable that said knowledgehase can be easily updated by a camera manufacturer and. to some extent. modified by an end—user. Thus various embodiments would store. or allow updating ofthe knowledgebase from (i) a compact tlash or other memory card; (ii) a USB link to a personal computer; (iii) a network connection for a networked/wireless camera and (iv) from a mobile phone network for a camera which incorporates the Tuitctionality of a mobile phone. In other alternative embodiments. where the camera is networked. the knowledgebase may reside on a remote server and may respond to requests from the camera for the resolving olia certain set of correction interdependeneies.

Now once the corrective image processing chain has been determined it is applied to the image 412 and a number of sanity checks are applied 412 to ensure that the image quality is not degraded by the correction process 416. If these tests fail then it may be that the determined interdependencies were marginal or that an alternative image processing strategy is still available 418. If this is so then the image processing chain is modified 420 and corrective image processing is reapplied 412. This loop may continue until all alternative image processing chains have been exhausted. It is further remarked that the entire image processing chain may not be applied each time. For example, if the differences between image processing chains is a single filter then a temporary copy of the input image to that filter is kept and said filter is simply reapplied with different parameter settings. If. however step 418 determines that all corrective measures have been tried it will next move to step -122 which determines if filter chain adaption is possible. Now returning to step 416, if the corrective image processing is applied successfully then the image is passed on to the redeye filter 90.

Fig 4(b) describes an alternative embodiment of the correction analyzer 135-2 which determines if filter chain adaption is possible and then modifies the redeye filter appropriately.

Initially the image characteristics list is loaded 401 and for each characteristic a set of filters which require adaption is determined 452. This is achieved through referencing the external knowledgebase 135-3 and the comments and discussion provided in the coiitext of the image correction knowledgebase 135-4 apply equally here.

Now once the filter lists for each image characteristic have been determined the correction analyzer must determine which filters overlap a plurality of image characteristics -154 and. additionally determine if there are conflicts between the filter adaptions required for each of the plurality of image characteristics 456. lf such conflicts exist the correction analyzer must next decide if they can be resolved 460. To provide a simple illustrative example we consider two image cliaracteristics which both require an adaption ofthe threshold of the main redness filter in order to compensate for the measured non-optimallity of each. If the first characteristic requires a lowering of the redness threshold by. say. 10% and the second characteristic requires a lowering of the same threshold by, say 15% then the correction analyzer must next determine from the knowledgebase the result of compensating for the first characteristic with a lowered threshold of % rather than the initially requested 10%. Such an adjustment will normal be an inclusive one and the correction analyzer may determine that the conflict can be resolved by adapting the threshold of the main redness filter to 15%. However it might also determine that the additional % reduction in said threshold will lead to an unacceptable increase in false positives during the redeye filtering process and that this particular conflict cannot be simply resolved. lfsueh lilter conflicts cannot be simply resolved an alternative is to determine if they are separable -"166. lf they are separable that implies that two distinct redeye filter processes can be run with different filter chains and the results of the two detection processes can be merged prior [J or to correcting the defects. In the case of the example provided above this implies that one detection process would be run to compensate for a first image characteristic with a threshold of % and a threshold of l5%. second detection process will be run for the second image characteristic with a The results of the two detection processes will then be combined in either an exclusive or an inclusive manner depending on the separability determination obtained from the sublilter knowledgebase 135-3.

Returning to step 460. we see that if filter conflicts can be resolved. the correction analyzer will prepare a single filter chain parameter list 462 which will then be loaded 46-! to the filter chain adapter 99 of the redeye filter 90 illustrated in Fig 3. Alternatively. if filter conflicts cannot be resolved. but are determined to be separable 466 the correction analyzer prepares a number of parameter lists 468 for the filter chain adapter which are then loaded 464 as in the previous case. The redeye filter is then applied 90.

Ilowcver. if filter conflicts cannot be resolved and are not separable the correction analyzer will then make a determination if image processing compensation might be possible 422. lf so then the image processing compensation workllow of Fig 4(a) may be additionally employed 400. If it is determined that image processing compensation is not possible then the correction analyzer will disable the redeye filter for this image 220. and provide a user indication to that effect 140 after which it will pass control back to the main in-camera application 224.

Fig 4(c) describes the workflow of the image analysis prefilter 130 illustrated in Fig l(b).

This performs an image processing analysis ofat least one image characteristic according to at least one of a plurality of image processing techniques. Preferably. the output of this analysis should be a simple measure of goodness ofthe analyzed image characteristic. For the purposes of an exemplary discussion we suppose that said measure is a percentage of the optimum for said characteristic. Thus l00% represents perfect quality for the measured image characteristic: values above 95% represent negligible image distortions/imperfections in said characteristic; values above 85% represent noticeable. but easily correctable distortions/imperfections and Values above 60% represent major distortions/imperfections which require major image processing to correct the image characteristic. Values below 60% imply that the image is too badly distorted to be correctable.

[Q U1 The first step in this workflow is to load or. if it is already loaded in memory. to access the image to be analyzed. The analysis prefilter next analyzes a first characteristic of‘ said image 482 and determines a measure of goodness. Now if said characteristic is above a first threshold (95%) 486 then it is marked as not requiring corrective measures 487 in the characteristic list. ll‘ it is below said first threshold, but above a second threshold (85%) 488 then it is marked as requiring secondary corrective measures 489. ll‘ it is below said second threshold. but above a third threshold (60%) 490 then it is marked as requiring primary corrective measures 491 and it‘ below said third threshold 492 it is marked as uncorreetable 493. Now it is remarked that for some embodiments of‘ the present invention which combine corrective image processing with filter chain adaption there may be two distinct sets of‘ thresholds. one relating to the correctability using image processing techniques and the second relating to the degree of‘ compensation possible using filter chain adaption. We further remark that for image compensation through filter chain adaption that certain filters may advantageously scale their input parameters directly according to the measure of goodness of certain image characteristics. As an illustrative example consider the redness threshold of the main color filter which. over certain ranges of‘ values. may be scaled directly according to a measure of excessive "redness" in the color balance of‘ a non- optimally acquired image. Thus, the image characteristic list may additionally include the raw measure of goodness of each image characteristic. In an alternative embodiment only the raw measure of‘ goodness will be exported from the image analysis prefilter 130 and the threshold based determining of Fig 4(c) will be performed within the correction analyzer 135-2 in which case threshold values may be determined from the image correction knowledgebasc 135-4.

Rcturning to 493 we note that images of‘ such poor quality may require :1 second image acquisition process to be initiated and so it is implicit in 493 that for certain embodiments of‘ the present invention it maybe desirable that an alarm/interrupt indication is sent to the main camera application.

Now the main loop continues by determining if the currently analyzed characteristic is the last image characteristic to be analyzed 496. If‘ not it returns to analyzing the next image characteristic 482. If‘ it is the last characteristic it then passes the image characteristics list to the image compensation prefilter 494 and returns control to the main camera application 224. ll should be remarked that in certain embodiments that a plurality of image characteristics may be '7’? grouped together and analyzed concurrently. rather than on a one—by—onc basis. This may be preferable if several image characteristics have significant overlap in the image processing steps required to evaluate them. It may also be preferable where a hardware co—proccssor or DSP unit is available as part of the camera hardware and it is desired to batch run or parallelize the computing ofimage characteristics on such hardware subsystems.

A third principle embodiment of the present invention has already been briefly described.

This is the use ofa global pixel-level transformation of the image within the redeye lilter itself and relies on the corrective image processing. as determined by the correction analyzer 135-2. being implementable as a global pixel—level transformation of the image. Those skilled in the art will realize that such a requirement implies that certain of the image analyzer elements which comprise the image analysis prefilter 130 are not relevant to this embodiment. For example dust analysis. object/region analysis. noise analysis and certain forms of image blur cannot be corrected by such transformations. However many other image characteristics are susceptible to such transformations. Further. we remark that this alternative embodiment may be combined with the other two principle embodiments of the invention to compliment each other.

In Fig 5(a) we illustrate an exemplary embodiment ofthe red pixel locating and red region segmenting workflow which occurs within the redeye filter as steps 92-1 and ‘)2-2. This workflow has been modified to incorporate a global pixel—level transformation 92-0 ofthe image as an integral element ofthe color determining and region grouping steps ofthe redeye filter. It is implicit in this embodiment that the correction analyzer has determined that a global pixel level transformation can achieve the required image compensation. The image to be processed by the redeye filter is first loaded 502 and the labeling l.UT for the region grouping process in initialized 504. Next the current pixel and pixel neighbourhoods are initialized 506.

Fig 5(b) shows a diagrammatic representation ofa 4-pixel neighborhood 562. shaded light gray in the figure and containing the three upper pixels and the pixel to the left of the current pixel 560, shaded dark gray in the figure. This 4—pixel neighborhood is used in the labeling algorithm of this exemplary embodiment. A look-up table. l.l.'T. is defined to hold correspondence labels.

Returning to step 506 we see that after initialization is completed the next step for the workflow of Fig 5(a) is to begin a recursive iteration through all the pixels of an image in a Ix) Ra.) raster-scan from top—left to bottom—right. The first operation on each pixel is to appl_\ the global pixel transformation 508. It is assumed that the loaded image is an RGB bitmap and the global pixel t1‘ai1sFo1‘1natio1i is ofthe form: P(R.G.B) -—> P(R'.(j'.B'). where the red. green and blue values ol‘ the current pixel. l’(R.G.B) are mapped to a shilted set ol‘ color space values, P(R'.G',B'). There are a number of advantages in performing this corrective transformation at the same time as the color determining and pixel grouping. In particular it is easier to optimize the computational performance ol‘ the algorithm which is important tor in- camera implementations. Following step 508 the workflow next determines if the current pixel satisfies membership criteria for a candidate redeye region 510. Essentially this implies that the current pixel has color properties which are compatible with an eye detect: this does not necessarily imply that the pixel is red as a range of other colors can be associated with llash eye defects. If the current pixel satisfies membership criteria for a segment 510. i.e.. it‘ it is stifficiently "red", then the algorithm checks for other "red" pixels in the 4—pixel neighborhood , If there are no other "red" pixels. then the current pixel is assigned membership ol‘ the current label 530. The LUT is then updated 532 and the current label value is incremented 534. it there are other "red" pixels in the 4—pixel neighborhood then the current pixel is given membership in the segment with the lowest label value 514 and the LUT is updated accordingly 516. Alter the current pixel has been labeled as part ofa "red" segment 512 or 530. or has been categorized as "non—red" during step 510, a test is then performed to determine ill it is the last pixel in the image 518. If the current pixel is the last pixel in the image then a final update otthe l,U'l' is performed 540. Otherwise the next image pixel is obtained by incrementing the current pixel pointer 520 and returning to step 508 and is processed in the same manner. Once the linal image pixel is processed and the final lrlll completed 540. all of the pixels with segment membership are sorted into a labeled—segment table of potential red—e_ve segments 542.

With regard to the exemplary details ofcorrective image processing 135-] which may be employed in the present invention we remark that a broad range of techniques exist for automatic or semi—automatic image correction and enhancement. For ease ofclisetission we can group these I J L/I into 6 main subcategories as follows: (i) Contrast Normalization and Image Sharpening. (ii) Image Color Adjustment and Tone Reproduction Scaling. (iii) Exposure Adjustment and Digital Fill Flash (iv) Brightness Adjustment with (‘olor Space Matching: Image /\uto—Gamma determination with Image Enhancement. (v) In-Camera Image Enhancement (vi) Face Based Image Fnhanccment All categories may be global correction or local region based Q) Contrast Normalization and Image Sharpening: US patent 6,421,468 to Ratnakar et al. disclose sharpening an image by transforming the image representation into a frequency—domain representation and by selectively applying scaling factors to certain frequency domain characteristics of an image. The modified frequency domain representation is then back-transformed into the spatial domain and provides a sharpened version of the original image. US patent 6,393,148 to Bhaskar discloses automatic contrast enhancement of an image by increasing the dynamic range of the tone levels within an image without causing distortion or shifts to the color map of said image.

Q1) Color Adjustment and Tone Scaling ofa Digital Image: US patent application 2002/0105662 to Patton et al. discloses modifying a portion of an image in accordance with colormetrie parameters. l\/lore particularly it discloses the steps of (i) identifying a region representing skin tone in an image; (ii) displaying a plurality of renderings for said skin tone: (iii) allowing a user to select one of said renderings and (iv) modifying the skin tone regions in the images in accordance with the rendering of said skin tone selected by the user. US patent 6,438,264 to Gallagher et al. discloses compensating image color when adjusting the contrast of a digital color image including the steps of (i) receiving a tone scale function; (ii) calculating a local slope of the tone scale function for each pixel of the digital image; (iii) calculating a color saturation signal from the digital color image and (iv) adjusting IX) ‘J1 the color saturation signal for each pixel ofthe color image based on the local tone scale slope.

The image enhancements of Gallagher et al. are applied to the entire image and are based on a global tone scale function. Thus this technique may be implemented as a global pi.\'el—le\ cl color space transformation. US patent 6,249,315 to Holm teaches how a spatially blurred and sub- sampled version of an original image can be used to obtain statistical characteristics of a scene or original image. This information is combined with the tone reproduction curves and other characteristics of an output device or media to provide an enhancement strategy for optiinized output of a digital image. All of this processing can be performed automatically. although the Holm also allows for simple. intuitive manual adjustment by a user. (iii) Digital Fill Flashzand Post—Acquisition Expostii‘e Adjustment US patent application 2003/0052991 to Stavely et al. discloses simulating till tlash in digital photography. In Stavely a digital camera shoots a series of photographs of a scene at various focal distances. These pictures are subsequently analyzed to determine the distances to different objects in the scene. Then regions of these pictures have their brightness selectively adjusted based on the aforementioned distance calculations and are then combined to form a single. photographic image. US patent application 2001/0031142 to Whiteside is concerned V\iil1 a scene recognition method and a system using brightness and ranging mapping. It uses auto- ranging and brightness measurements to adjust image exposure to ensure that both baclxground and foreground objects are correctly illuminated in a digital image. Much ofthe earlier prior art is focused on the application of corrections and enhancement of the entire image. rather than on selected regions of an image and thus discuss the correction of image exposure and tone scale as opposed to fill flash. Example patents include US patent 6,473,199 to (iilman et al. which describes a method for correcting for exposure in a digital image and includes proiiding a plurality of exposure and tone scale correcting nonlinear transforms and selecting the appropriate nonlinear transform from the plurality of nonlinear transforms and transforming the digital image to produce a new digital image which is corrected for exposure and tone scale. liS patent ,991,456 to Rahman et al. describes a method of improving a digital image. lhe image is initially represented by digital data indexed to represent positions on a display. The digital data is indicative of an intensity Value li (x.y) for each position (x.y) in each i—th spectral band. The i\> ui intensity value for each position in each i-th spectral band is adjusted to generate an adjusted intensity value for each position in each i-th spectral band. Each surround function l‘n txpy) is uniquely sealed to improve an aspect ofthe digital image, e.g.. dynamic range compression. color constancy, and lightness rendition. For color images. a novel color restoration step is added to give the image true—to-life color that closely matches human observation.

However some of the earlier prior art does teach the concept of regional anal_vsis and regional adjustment of image intensity or exposure levels. US patent 5,818,975 to Goodwin et al. teaches area selective exposure adjustment. Goodwin describes how a digital image can have the dynamic range of its scene brightness reduced to suit the available dynamic brightness range of an output device by separating the scene into two regions — one with a high brightness range and one with a low brightness range. A brightness transform is derived for both regions to reduce the brightness of the first region and to boost the brightness of the second region. recombining both regions to reform an enhanced version of the original image for the output device. lhis technique is analogous to an early implementation of digital fill flash. Another example is US patent 5,724,456 to Boyack et al. which teaches brightness adjustment of images using digital scene analysis. Boyaek partitions the image into blocks and larger groups of blocks. known as sectors. lt then determines an average luminance block value. A dilference is determined between the max and min block Values for each sector. If this difference exceeds a pre—deter1nined threshold the sector is marked active. A histogram of weighted counts of active sectors against average luminance sector Values is plotted and the histogram is shifted to using a pre-determined criteria so that the average luminance sector values of interest will fall within a destination window corresponding to the tonal reproduction capability of a destination application or output device. ( iv) Brightness Adjustment; Color Space Matching; Auto—Gamma.

Another area of image enhancement in the prior art relates to brightness adjustment and color matching between color spaces. For example US patent 6,459,436 to Kumada et :11. describes transforming image date from device dependent color spaces to deviee—independent Lab color spaces and back again. Image data is initially captured in a color space representation which is dependent on the input device. This is subsequently converted into a device independent color space. Gamut mapping (hue restoration) is performed in the device independent color space and the image data may then be mapped back to a second device—dependent color space. US patent 6,268,939 to Klassen ct al. is also concerned correcting luminance and chrominanee data in digital color images. More specifically Klassen is concerned with optimi7ing the transformations between device dependent and device independent color spaces by applying subsampling ofthe luminance and chrominance data. Another patent in this category is ITS patent 6,192,149 to Eschback et al. which discloses improving the quality of a printed image by automatically determining the image gamma and then adjusting the gamma of a printer to correspond to that of the image. Although Esehback is concerned with enhancing the printed quality ofa digital image and not the digital image itself. ifdoes teach a means for automatically determining the gamma of a digital image. This information could be used to directly adjust image gamma. or used as a basis for applying other enhancements to the original digital image.

US patent 6,101,271 to Yamashita et al. discloses implementing a gradation correction to an RGB image signal which allows image brightness to be adjusted without affecting the image hue and saturation. (v) ln—Camera Image Enhancement US patent 6,516,154 to Parulski et al. discloses suggesting improventcnts to a digital image after it has been captured by a camera. The user may crop. re—size or adjust color balance before saving a picture; alternatively the user may choose to re—take a picture tising different settings on the camera. The suggestion of improvements is tnade by the camera user—interfacc.

Ilowever Parulski does not teach the use of image analysis and corrective image processing to automatically initiate in-camera corrective actions upon an acquired digital image. (vii) Face—Based Image Enhancement In US patent application 20020172419, Lin et al.. discloses automatically improxing the appearance of faces in images based on automatically detecting such images in the digital image. l.in describes modification of lightness contrast and color levels of the image to produce better results.

Claims (5)

CLAIMS:

1. A method for red-eye detection in an acquired digital image comprising the steps ol‘: a) acquiring a tirst image; b) analysing the first acquired image to provide a plurality ot‘characteristics indicative of image quality; c) determining if one or more corrective processes can be beneficially applied to said tirst acquired image according to said characteristics; d) applying any such corrective processes to said first acquired image; and e) detecting red-cyc defects in a second acquired image using said corrected tirst acquired image.

2. A method according to claim 1 wherein said detecting step comprises applying a chain of one or more red-eye filters to said first acquired image; and further comprising the steps ol‘. prior to said detecting step: D determining if said red-cyc filter chain can be adapted in accordance with said plurality of characteristics; and g) adapting said red-cyc filter chain accordingly.

3. A method according to claim 2 in which said step of adapting comprises proxriding an altered set of parameters for one or more filters ofsaid filter chain.

4. A computer—readable storage medium containing a set ofiitstrtictions nhich when executed on a digital image processing device perform the steps oilclaim l.

5. A digital image processing device arranged to perform the steps of claim 1.