An exposure fusion scheme for differently exposed images with moving objects is presented here. Block diagram of the system is as illustrated in Figure 1. This system involves a ghost removal algorithm in a low dynamic range domain and a selectively detail-enhanced exposure fusion algorithm.
The input to the system would be three differently exposed images by varying exposure time we get over exposed ,medium exposed and under exposed images. Image is described as overexposed (bright) when it has a loss of highlight details ,that is , bright area of the image are ‘washed out’or effectively all white known as blown out highlights. Image is described as underexposed(dark) when it has a loss of shadow details, that is, dark regions are indistinguishable from black known as ‘blocked up shadows’.
The input images undergo pre-processing. Pre-processing step involves resizing images and conversion of the images form RGB color space to YCbCr color space .RGB represents color as red ,green and blue . YCbCr represents Color as brightness and two color difference signals Y is brightness (luma) which is very similar to grayscale version of original image .Cr is red minus luma (R-Y) and Cb is blue minus luma (B-Y) .luma component represents brightness of pixel and chroma component represents color of pixels. color space conversion has become essential part of image processing and transmission. Images and videos are stored in rgb color space .transmission of these images and videos is not feasible as their bandwidth requirement is more. To overcome this problem and to reduce bandwidth requirements of images in rgb color space are converted to other color space such as yCbCr and then can be transmitted . Y plane is extracted from all three images and are passed to ghost removal model.
Ghost removal model.
The ghost removal algorithm is composed of two modules: a detection module and a correction module. Initially, in all the input images, non-consistent pixels are detected by using the detection module. In detection module we select the reference image and threshold value .middle image is selected as initial reference image according to overall exposedness of luminance components in all input images . for every image except the refrence Further the correction module is used to correct all the non-consistent pixels. This results in consistent pixels in all the corrected images. Contrast of Y planes extracted from all three images is measured. In order to measure contrast, Laplacian filter is applied and absolute value of the filter response is computed. The results from contrast measurement algorithm and ghost removal are used to blend the pixel intensities. The outcome obtained after blending is passed to selectively detail-enhanced fusion stage. This algorithm is introduced to enhance fine details in all regions of the image. The result is a perfectly fused image.
...(download the rest of the essay above)