Outdoor Image Enhancement:Increasing Visibility Under Extreme Haze and Lighting Condition

Transcription

1 Outdoor Image Enhanement:Inreasing Visibility Under Extreme Haze and Lighting Condition Deepak Kumar Naik Dept. of Eletronis & Teleommuniation C. V. Raman Polytehni Bhubaneswar, India Deepak Kumar Rout Image Analysis & Computer Vision Lab. Dept. of Eletronis & Teleommuniation Engineering C. V. Raman College of Engineering Bhubaneswar, India Abstrat This paper addresses the problem of image enhanement thereby enhanement of sene visibility in outdoor images. Visibility is a very important issue in ase of omputer based surveillane, rime analysis, driver assistane system design et. The most important hallenge related to visibility is the atmospheri haze and poor lighting. The problem beomes more hallenging if haze is too dense and lighting during night is extremely poor. In this paper an automati degradation detetion and restoration algorithm has been proposed, whih detets the type of degradation using the distribution of the sene, then uses the hybrid dark hannel prior based haze removal algorithm if the image is degraded due to atmospheri haze only, otherwise it omputes the image negative first and then uses the hybrid DCP to resolve the problem. The algorithm has been tested in many situations and the results obtained are satisfatory as omparison to the existing algorithms. Keywords- dark hannel prior, histogram equalization, hybrid DCP, poor lighting, image dehazing, auto-haze detetion. I. INTRODUCTION The image proessing is the vast emerging field in the era of tehnology of mahine vision, mahine intelligene and automation for real time proessing or the post proessing of the image aptured in different atmospheri onditions. The image aptured in the outdoor sene are highly degraded due to the poor lighting ondition or over lighting ondition or due to the presene of different suspension partile like the water droplets or dust partiles. So due to these partiles the irradiane oming from the objet is sattered or absorbed. And hene the phenomena of haze, smoke and fog ours. So the images are degraded and the olor, ontrast are shifted from its original irradiane at the time of apture of the image. If the image is de-hazed then it an be proessed and an be implemented it the field of the omputer vision analysis and robotis. The haze removal is very essential in the field of image proessing beause the different omputer vision algorithm assumes the input image as the original sene radiane or sene refletane. But in most outdoor proessing the images are degraded due to hazy, hene the input image is hazy image not the original radiane. If the haze an be removed then the sene will have proper brightness, ontrast and the information ontents in the image will be high. The haze removal proess is very ompliated beause the haze depends upon the unknown depth of the objet in the sene. The seond problem whih has been onsidered in this paper is enhanement of image when it is aptured under night ondition. In this ase the objet is rarely visible and hene the aptured image has less amount of information. So the extration of information from suh images is diffiult. In the other hand the appliation of infrared amera often fails in the detetion of objet and bakground when both have same temperature. The onventional method for the night image enhanement is to invert the image and then enhane it with histogram equalization. But these methods are sensitive to the deviation of olor hannel from its original olor due to global enhanement tehnique. So output images may be unreal in pereption. In the field of haze removal different methods are present whih basially built on the priority of strong assumptions. Tan [6] and Fatal [5] proposed different algorithms based upon the priority and their algorithm an enhane the hazy images appreiably. But pratially in some physial onditions the assumed models may fail and may provide physially invalid output results. These models may not enhane the extreme hazy images. Kiming He [1] also proposed algorithm for single image haze removal using the dark hannel prior method whih is also based upon a strong priority and statistis of the outdoor haze-free images. The dark hannel prior method is very useful to dehaze the hazy images but fails in some speifi onditions, to be more speifi, it fails to enhane the sky regions where the sunlight is very influential. So in outdoor image ase the dark hannel prior method is ineffiient to enhane the entire sene. In this paper, we propose a highly effiient and integrated algorithm to enhane both hazy outdoor images and the low light images i.e. taken in the night ondition. The proposed algorithm integrates both the algorithms for hazy image as well as the night images. A new method is proposed where we integrate histogram equalization with the dark hannel prior method so as to de-haze the image as well as to /14/$ IEEE 1081

2 inrease the ontrast of the image. This proposed method also eliminates the problem arising in the sky region enhanement in the method proposed by Kiming He [1]. This proposed method utilizes the ore de-haze algorithm and histogram equalization to enhane the outdoor images in extremely poor visibility ondition. II. ENHANCEMENT OF HAZY AND LOW LIGHTING IMAGES A. Algorithm to Detet the Type of Images: In order to detet, whether the image is degraded by haze during day light or it is degraded due to insuffiient lighting during night time, the following algorithm has been used. Plot the histogram of the input image Determine the mean of the intensity distribution(let it be m ) Compare this mean with the pre-defined threshold to deide, whether the image is degraded by haze during day light or it is degraded due to insuffiient lighting during night time. Take a threshold(say ) If m > Then the image is degraded by haze else the image is degraded due to insuffiient lighting of sene during night. One the type of degradation is deided, then the orresponding enhanement algorithms an be used to enhane the sene visibility. The flow diagram of the algorithm to detet the input images is shown in Fig. 1 omponents present in a pixel of the image. But the sky region have the high intensity as ompared to the objet whih we apture. So the non-sky region of the image have the lower intensity of the pixel value and hene the image may be onsidered as the hazy image due to the deviation of olor of the different objets present in the image. In [1], Kiming He proposed the ore de-haze model on the basis of the outdoor image statistis analysis. He observed that when the image is a haze-free, then the minimum intensity of the different pixel value is very low in numbers. But when the image is a hazy one then the number of pixel having intensity low are high. So aording to the dark hannel prior method the amount of minimum intensity value an be desribed as a funtion and denoted as J dark ( and it an be represented as: dark J ( = min min J ( y) (1) ( { r, g, b} Where J is a olor hannel of J and ( is a loal path antered at x. It is the outome of two minimum operator i.e. min { r, g, b} is minimum operator that operates on eah pixel and min ( is the minimum filter that operates on the path of size Ω(. so aording to the onept of dark hannel prior method J dark ( has to be defined for hazy images. If the image is a hazy image then the intensity of J s dark hannel will be low and it tends to zero value: J dark ( 0 (2) This expression is alled as dark hannel prior. This priority is utilized to de-haze the images. The haze image an be desribed by using [2], [5], [8], [10], [11]: I( = J ( + A(1 ) (3) Where I is the observed intensity of the sene, J is the original sene radiane, A is the atmospheri light and t is the medium transmission oeffiient whih desribes the amount of light propagate to the amera without sattering. The original sene radiane J( an be alulated from the other parameter if we an estimate the and A properly. The I( is our observed intensity. And hene a haze free image an be estimated as: Fig. 1 Flow diagram to detet the input images B. Dark Channel Prior Method with Histogram equalization: Image Dehazing The images aptured in the outdoor hazy ondition suffer in the deviation of different olor hannel of the RGB I ( A J ( = + A (4) And this is the expression of haze free images derived from (3). Hene the dark hannel prior method an be implemented to suessively reover the original image radiane from the observed degraded hazy images. In the next step the transmission oeffiient has to be derived by applying the IEEE International Advane Computing Conferene (IACC)

3 dark hannel prior tehnique. The an be alulated from the haze model (3). From equation (3) we have: I J ( = + 1. (5) A A Here eah olor hannel is normalized independently with the atmospheri light. Now on both side of (5) we have to alulate the dark hannel prior. So the two minimum operator has to be applied on both side. I ( y) J ( y) min min = min min + 1. ( A ( A (6) The transmission oeffiient t ( is taken outside the minimum filter as this an be onsidered as onstant in a path. As the sene radiane J is a haze-free image, the dark hannel of J is lose to zero due to the dark hannel prior: dark J ( = min min J ( y) = 0 (7) As ( ( ). A is always positive, this leads to: J ( y) min min x ) A ( = 0. Now putting the values of equation (8) in equation (6), we an eliminate the multipliative term and estimate the transmission t ( simply by: I ( y) t ( = 1 min min. (9) ( A I ( y) Here min min is the dark hannel of the ( x ) A I ( y) normalized hazy image. It diretly provides the A estimation of the transmission oeffiient. When the haze of an image is ompletely removed then it may look like unreal and physially invalid or the depth of the image may be loosed. So to keep some aerial perspetive or the depth of the image it is very muh required to keep some amount of haze so that the image will be visually pleasant and ontains high amount of visual information. To keep some haze a parameter ω ( 0 < ω 1) is introdued in the transmission oeffiient: I ( y) = 1 ω min min (10) ( A The value of ω depends upon different input onditions. Its value vary from sene to sene. But for the hazy images its (8) value an vary in between 0.75 to 0.95 with respet to the amount of haziness of the image. If the quantity of haze is very high then and ω is taken 0.95 then the transmission of different olour hannel is affeted and the shift of olour in the original radiane is deviated. Often the hazy images beome darker or mostly bluish in olour. So the seletion of this weight parameter ω undergoes a haze based algorithm whih alulate whether the ω value will be higher (ω =0.75) or lower (ω =0.95). So it is now adaptive to hoose an appropriate value of weightω. Now by putting (10) in (4) the original radiane of the sene an be alulated and hene a haze-free image an be reovered i.e. I ( A J ( = + A (11) The very onventional method to inrease the visibility of the image is to take the histogram equalization of the degraded or dull images. But when this equalization proess is applied, the ontrast level inrease to a ertain amount but the haze also inrease simultaneously. At the same time the dark hannel prior an remove the haze and give proper radiane of the sene. So a new algorithm has been proposed whih integrates both dark hannel prior method and histogram equalization method. But between dark hannel prior method and histogram equalization method, whih has to be proessed first it has to be taken arefully by using ertain algorithm. Otherwise this algorithm will not be adaptive and in ertain ase the images willl be degraded entirley. By taking different types of haze image it observed that when the amount of haze is extremely high then if histogram equalization is applied, then image will be hazier. So in this ase the dark hannel prior will be applied first and then the histogram equalization proess will be very effetive. And the result will be haze free as well as high ontrast image. In the seond observation when the amount of haze is little lower then first equalization will be applied and then dark hannel prior will be applied and output will be more pleasant than the dark hannel prior image alone. When histogram equalization proess is applied to the reovered image obtained in equation (11), the enhaned image an be represented as: k k n k j 1 J = = = = equa ( T ( rk ) pr ( rj ) n j j= 0 j= 0 M * N M * N j= 0 (12) The equation (12) is the equalized enhaned image, where J equa ( is the equalized output of the image J ( T ( r ), k is the transformation funtion ating upon the input image r p ( ) intensity k, r r j is the probability of ourrene of the r n image intensity k, j is the number of pixels that have the 2014 IEEE International Advane Computing Conferene (IACC) 1083

4 rj intensity and M*N is the total number of pixel present in the image.dark Channel Prior. Fig. 2 Flow diagram to detet amount of haziness present in the input hazy image. So by taking this observation we have defined the algorithm. We have taken a threshold value (let ). When amount of haze is greater than, then the image is said to be extreme hazy image. And when amount of haze is less than, then it is alled less hazy image as shown in Fig. 2. Now we an apply our proposed algorithm. C. Enhanement of Night Images: Inreasing Visibility in Night The low-light images or the night images ontains very less amount of visual information. So to extrat information we have to go for image negative. But when images are inverted then inverted images has very similar statistis as the hazy images. In both type of images i.e. inverted night images and hazy images have the same harateristis (as desribed in [1]) that the intensity of bakground pixel is always high in all olor hannel and the intensity of objet is very less in at least one olor hannel. So by using the statistis [2], the dark hannel prior method an be implemented to the low-light images or night images after the inversion of the input images. So the inversion proess an be represented as: R ( = 255 I ( (13) where is the olor hannel i.e. RGB, I ( is the intensity of the pixel of an input low light image. Here R ( is the same intensity of the inverted image R. Hene the haze model for the inverted night / low-light image an be represented as: R( = J ( + A(1 ) (14) where A is the atmospheri light and J( is the original intensity of the objet or sene. Here is the transmission oeffiient as desribed in the (10). So now for night image or low-light image the transmission oeffiient an be expressed as: R ( y) = 1 ω min min (15) ( A where ω is the weight parameter and it is taken lower than that of ω taken for the hazy images. Its typial value is taken as 0.85 in this paper. The value of atmospheri light A an be alulated from the pixel of the input image. The maximum value of all the olor hannel i.e. RGB is alulated and the maximum values alulated are taken as (A). The reovered image is represented as: R( A J ( = + A (16) After the reovery of the value J, then the again it has to be inverted so as to get the enhaned output images.so we have proposed an integrated algorithm whih will detet as well as proess both night images and day hazy images. This proposed algorithm also adaptively hoose the value of weight parameterω. The formulated algorithm is shown in the Fig.3. Fig. 3 Flow diagram of the module to determine the input image and then proess with histogram based dark hannel prior method. III. EXPERIMENTAL RESULTS The hazy and low-light images are proessed by the system having the speifiation: Intel(R) Core(TM) i Hz 2.40 Hz, Installed Memory(RAM): 4.00GB, System Type: x64-based proessor, OS- Ubuntu IEEE International Advane Computing Conferene (IACC)

5 The images taken in the experiment have the resolution maximum up to 500x500. To proess the input images the system takes maximum up to 1 Se. In the Fig. 4 the images are taken in the night ondition, so these are desribed as low light images. Hene they are first inverted and then dark hannel prior algorithm is used. The resultant image is then inverted to get the enhaned images. In the Fig. 5, the transmission map have been ompared between the input image and the proessed output image by our method. In the transmission map of proessed image maximum amount of information has been transmitted. In Fig. 6, it desribes about the failure of He. Et al. s dark hannel prior results in the sky region. So here a qualitative omparison has been introdued. In Fig. 6(b) a red olor arrow has been shown where the He. Et al. s dark hannel prior results fails and it beomes bluish. In Fig. 6 () our result is shown where in sky region, our proposed algorithm is valid with proper enhanement. In Fig. 7 the results are ompared among the histogram equalization result, dark hannel prior results and our proposed method results and our results are better in qualitative view. IV. CONCLUSION In this paper we have proposed a very simple and effiient algorithm whih will first detet and then proess with proper algorithm for different atmospheri ondition. This proposed work also eliminates the drawbaks arising in other existing methods. But it is observed that the weight parameter ω widely varies from sene to sene. So it beomes very important parameter to deal with the haze image. Therefore our proposed model fails to make the parameter ω adaptive for all weather ondition. We leave this for future work so that a more effiient algorithm an be developed to remove the haze. REFERENCES [1] Kaiming He, Jian Sun, and Xiaoou Tang. Single Image Haze Removal Using Dark Channel Prior. IEEE Transation on Pattern Analysis and Mahine Intelligene, VOL. 33, NO. 12, Deember [2] Xuan Dong, Guan Wang, Yi Pang, Weixin Li. Fast effiient algorithm for enhanement of low lighting video. Multimedia and Expo (ICME), 2011 IEEE International Conferene Barelona, pp. 1-6, July [3] S.G. Narasimhan and S.K. Nayar, Vision and the Atmosphere, Int l J. Computer Vision, vol. 48, pp , [4] S.G. Narasimhan and S.K. Nayar, Chromati Framework for Vision in Bad Weather, Pro. IEEE Conf. Computer Vision and Pattern Reognition, vol. 1, pp , June [5] R. Fattal. Single Image Dehazing, in ACM SIGGRAPH 08, Los Angeles, CA, Aug. 2008, pp [6] R. Tan. Visibility in Bad Weather from A Single Image, in Pro. IEEE Conf. Computer Vision and Pattern Reognition., Anhorage, Alaska, pp. 1-8, Jun [7] K. He, J. Sun, and X. Tang. Single Image Haze Removal Using Dark Channel Prior, in Pro. IEEE Conf. Computer Vision and Pattern Reognition. Miami, FL, pp , Jun [8] H. Koshmieder, Theorie der Horizontalen Sihtweite, Beitr. Phys. Freien Atm., vol. 12, pp , [9] Y.Y. Shehner, S.G. Narasimhan, and S.K. Nayar, Instant Dehazing of Images Using Polarization, Pro. IEEE Conf. Computer Vision and Pattern Reognition, vol. 1, pp , [10] S. Shwartz, E. Namer, and Y.Y. Shehner, Blind Haze Separation, Pro. IEEE Conf. Computer Vision and Pattern Reognition, vol. 2, pp , [11] S.G. Narasimhan and S.K. Nayar, Contrast Restoration of Weather Degraded Images, IEEE Trans. Pattern Analysis and Mahine Intelligene, vol. 25, no. 6 pp , June [12] J. Kopf, B. Neubert, B. Chen, M. Cohen, D. Cohen-Or, O. Deussen, M. Uyttendaele, and D. Lishinski, Deep Photo: Model-Based Photograph Enhanement and Viewing, ACM Trans. Graphis, vol. 27, no. 5, pp. 116:1-116:10, [13] S.G. Narasimhan and S.K. Nayar, Interative Deweathering of an Image Using Physial Models, Pro. IEEE Workshop Color and Photometri Methods in Computer Vision, in Conjuntion with IEEE Int l Conf. Computer Vision, Ot [14] H. Ngo, L. Tao, M. Zhang, A. Livingston, and V. Asari. A Visibility Improvement System for Low Vision Drivers by Nonlinear Enhanement of Fused Visible and Infrared Video, in Pro. IEEE Conf. Computer Vision and Pattern Reognition., San Diego, CA,pp.25,, Jun [15] I. Omer and M. Werman, Color Lines: Image Speifi Color Representation, Pro. IEEE Conf. Computer Vision and Pattern Reognition, vol. 2, pp , June [16] E. Hsu, T. Mertens, S. Paris, S. Avidan, and F. Durand, Light Mixture Estimation for Spatially Varying White Balane, Pro. ACM SIGGRAPH 08, [17] A.J. Preetham, P. Shirley, and B. Smits, A Pratial Analyti Model for Daylight, Pro. ACM SIGGRAPH 99, [18] P. Chavez, An Improved Dark-Objet Subtration Tehnique for Atmospheri Sattering Corretion of Multispetral Data, Remote Sensing of Environment, vol. 24, pp , [19] R. Lim, T. Bretshneider. Autonomous Monitoring of Fire-related Haze from Spae, in Conf. Imaging Siene, Systems and Tehnology, Las Vegas, Nevada, Jun. 2004, pp Fig. 4 (a): Input low light image, (b): Inverted image, (): DCP of inverted image, (d): Our enhaned result 2014 IEEE International Advane Computing Conferene (IACC) 1085

6 (a) (b) () (d) Fig. 5 (a): Input hazy image,(b): Transmission map of (a),(): Transmission map after our method,(d): Our enhaned result (a) (b) () Fig. 6 (a): Input hazy image,(b): Result obtained using He et. al. method [1],(): Result obtained by our method (a) (b) () (d) Fig. 7 (a): Input hazy image,(b): Histogram equalized image, (): DCP result,(d): Our enhaned result IEEE International Advane Computing Conferene (IACC)