Shadow Remover Image Shadow Removal Based on Illumination Recovering Optimization（word版本） - 图文

Fig. 10. Color transfer. (a) Sample images, (b) input target images, (c) color transfer results, where color of the red box in the sample image is transferred to the target image.

(a) (b) (c)

Fig. 12. Shadow removal results. (a) Input images, (b) results of [8], (c) our results.

from green to yellow.

Fig. 11. Color transfer. (a) Target color image, (b) input image, (c) color transfer result with gx = 0.2, (d) color transfer result with gx = 0.4, (e) color transfer result with gx = 0.7, (f) color transfer result with gx = 0.9.

recovering: I^dlte = (fxLd + La)Rx = h， where fx is the normalized light attenuation function, and 0 < fx < 1. The user can develop more sophisticated attenuation function to produce more physically realistic results.

2) Color Transfer: Color and tone transfer is a popular topic in computer graphics [20], [37]. Our illumination recovering operator can also work on color transfer. Given the sample image, we can transfer its color to the target image. More specially, to transfer the color information of the sample image to the target areas, we apply our illumination recovering operator on these areas. The color transfer result for the target result can be written as: iedlt = (t + 1)Ix, where Ix denotes the RGB components of the original image. Fig. 10 shows two examples for color transfer.

We can also simulate the color transfer processing in animated style. We add a color transfer parameter gx to the color transfer equation, and gx is assigned by user according to different applications. It also can be set as a function. Then, the color transfer process can be expressed as: iedlt = (gxt +1)Ix. As shown in Fig. 11, there are some color transfer results by setting different values for gx , which can simulate the process for grass

VII. EXPERIMENTS AND DISCUSSIONS

In this section, we perform our shadow removal method on a variety of shadow images to illustrate the effectiveness of the proposed approach. We also present comparisons with the most related shadow removal methods. The limitations of the proposed method are also given. All our results are implemented using C++ on a machine equipped with Pentium(R) Dual-Core CPU E5200@2.50GHz with 2GB RAM.

In Fig. 12, we compare our method with the approach [8]. Guo et al. [8] segment the input image using texton histogram and SVM, but the divided regions are irregular. Large regions may contain several different kinds of colors and textures, which leads to calculation error for the ratio between direct light and environment light. This may result in an unsatisfactory shadow-free result, as shown in Fig. 12(b). Our method introduces adaptive overlapped patches which alleviate the problem of inaccurate matching. Meanwhile, we calculate different ratio values between direct illumination and envi-ronment illumination for different patches, but Guo et al. [8] uses the paired regions to compute a fixed ratio value for an image. The fixed ratio is impropriate for the shadow images with complex textures, which makes the shadow-free result unnatural, such as the result in Fig. 12(b).

In Fig. 13, we present shadow removal results for boundary treatment. As illustrated in Fig. 7, the texture details under the shadow boundaries may be blurred. We apply the texture and illumination optimization to recover the information of the shadow boundaries, and also present comparison results with previous methods. Xiao et al. [5] processes the shadow boundary using alpha matte interpolation. As the interpolation process depends on the original texture details at the shadow boundary regions, this method cannot process complex shadow boundaries with loss of details. Shor and Lischinski [10] utilizes graph-cut based texture technique to repair the shadow boundaries. This method works well for image with rich texture, otherwise, the results will not be satisfied. The method of [1] needs to locate the boundary precisely, and inaccurate shadow edges lead to fuzzy transition from interior shadow regions to nonshadow regions. Compared with aforementioned methods, our results are more visually natural and consistent with the surrounding content.

⑻ （b) (C) (d) (e)

Fig. 13. Boundary processing results. (a) Input images, (b) results of [1], (c) results of [10], (d) results of [5], (e) our results.

Fig. 14. Nonuniform shadow removal result. (a) Input image, (b) result of [1], (c) result of [10], (d) results of [5], (e) our result.

The results in Fig. 14 demonstrate that our method is capable of dealing with nonuniform shadows. Note that in these images, the shadow intensity vary among the umbra, penumbra, and lit regions. In Fig. 14, we compare our method with the related works [1], [5], [10] for handling nonuniform shadows. Since the relatively wide soft shadow on the shadow boundaries, the accurate shadow edges are difficult to be detected out. This makes Finlayson et al. [1] impossible to recover the illumination in shadow regions. Shor and Lischinski [10] use fixed parameters for whole shadow regions. But the illumination is variable for soft shadows, it make this method failed to deal with nonuniform shadows. The unsatisfied results in Fig. 14(d) demonstrate Xiao et al. [21] also cannot treat nonuniform shadows. By using adaptive texture patch matching, our method processes each small patch individually and remove the nonuniform shadows completely. In addition, the recovered illumination of nonuniform shadow regions is consistent with the surrounding environment.

In Fig. 15, we give shadow removal results for shadow images with complex textures, and compare our methodwith [1], [5], [8], [10]. These images contain multiple textures or materials. Shor and Lischinski [10] fails to take texture types into account in shadow regions. This make the wrong recovered color in shadow regions, especially for images with large color variation between differen textures. The shadow removal method [10] for simple texture image is not applicable to complex texture images. Xiao et al. [5] take the texture classification into account and divides the shadow areas into several parts. The independent processing for each part induces inconsistency between subregions, as shown in the fifth row in Fig. 15. Neither [1] nor [8] works well in removing shadows with complex texture or material content. Note that for the results of the sixth row in Fig. 15, even the shadow regions have a variety of texture types, our method can effectively recover the texture information in the shadow regions. In addition, the illumination on the shadow boundaries is also effectively recovered, and the results are consistent with the surrounding regions.

Our method can also be applied to aerial remote sensing shadow images. In Fig. 16, we compare our method to [5] and [6] on aerial remote sensing images. The texture

Fig. 15. Shadow removal results for complex textures. The first row: input images, the second row: results of [1], the third row: results of [10], the fourth row: results of [8], the fifth row: results of [5], the sixth row: our results, the patch size is 15 x 15, 20 x 20, 30 x 30, 21 x 21, respectively.

descriptor used in [5] is not illumination independent which The linear shadow-free algorithm [5] does not work well on is not suited for subregion matching between shadow regions aerial shadow image, as illustrated in Fig 16(b). Li et al. [6] and nonshadow regions. Meanwhile, as demonstrated by [38], removed the shadows of image using spatially adaptive high resolution aerial images often contain heavy noises. NL operators. To reduce the influences of abundant texture