Efficient Contrast Enhancement using Iterative Moving Average Filter

Abstract

Low contrast images occur from various reasons(unfavorable environmental condition, the inadequacy of the image capture device or lack of operator expertise). These factors result in under-utilization of the offered dynamic range; therefore, information of an image/video will be ambiguous and washed-out. Contrast enhancement targets to eliminate these problems. Contrast enhancement techniques can be broadly categorized into direct and indirect methods. Direct methods use defined contrast term as the criteria to measure the degree of improvement and to terminate the enhancement. Indirect methods exploit under-utilized regions of the dynamic range without a specific contrast term. Previous methods mostly fall into the indirect methods. Indirect methods can be further divided into several sub-groups: the frequency-based, histogram modification-based, and transform-based technique. Out of these three groups, the second group receives a lot of attention from researchers because of its straightforward and the intuitive process. Histogram modification techniques can be divided into three approaches; global, local, and unconventional approaches. Out of the three approaches, global approaches have been studied extensively because global approaches guarantee fast running time and are simpler and easier to implement than other approaches. Additionally, the state-of-the-art global approaches show better performance than local and unconventional approaches. The latest out of global approaches, Toet introduced a method that modifies a histogram using a power and log operation(PL). PL ideally modifies an input histogram through the quadratic optimization approach. Because the transformation function of PL reflects the shape of an input histogram, PL preserves pattern details of an input image during the contrast enhancement. However, PL may cause unanticipated artifacts that are invisible at the input image because the rest spikes of the modified histogram cause steep slopes of the transformation function, and these slopes may cause over-enhancement. On the other hand, Arriaga-Garcia’s method (bi-histogram equalization with adaptive sigmoid functions, BEASF) does not cause these artifacts. Because the transformation function of BEASF is estimated from sigmoid functions, the transformation function of BEASF is less influenced by the rest spikes of a histogram and smoother than the transformation function of PL. However, the transformation function of BEASF is difficult to reflect the shape of a histogram because the transformation function of BEASF is based on sigmoid functions. Therefore, BEASF may not sufficiently enhance low contrast pixels. Consequently, we formulate a hypothesis; the optimal transformation function should be estimated from the shape of an input histogram and have smooth curvatures to reduce steep slopes by the rest spikes. To verify the hypothesis, first we highlight limitations of traditional methods through the comparison of the difference between the transformation functions, and then we propose a method to estimate an optimal transformation function. The proposed method is constructed of a histogram modification and a transformation function smoothing. First, the proposed method modifies an input histogram using the power and log operation to remove big spikes of the input histogram. Next, the transformation function is calculated using the modified histogram. Finally, the proposed method smooths the transformation function through a regression model. The regression model performs a moving average filter to remove steep slopes of the transformation function and measures a terminal condition in every smoothing process. As a result, the proposed method estimates the optimal transformation function, that is less influenced by spikes of a histogram and reflects the shape information of a histogram. In the experiment, the proposed method is compared with 10 traditional methods by 3 contrast enhancement measurements and 1 noise measurement using famous 80 images. The proposed method demonstrates the best total rank score through quantitative measurements and the natural enhancement results through qualitative analysis. Also, the proposed method guarantees the similar complexity compared with the state-of-the-art methods. The proposed method may improve the recognition rate of users using TV, monitors and smart phones. Furthermore, the proposed method may decrease the error rate and increase efficiency of image processing algorithms.