Currently, the most advanced knowledge distillation models use a metric learning approach based on probability distributions. However, the correlation between supervised probability distributions is typically geometric and implicit, causing inefficiency and an inability to capture structural feature representations among different tasks. To overcome this problem, we propose a knowledge distillation loss using the robust sliced Wasserstein distance with geometric median (GMSW) to estimate the differences between the teacher and student representations. Due to the intuitive geometric properties of GMSW, the student model can effectively learn to align its produced hidden states from the teacher model, thereby establishing a robust correlation among implicit features. In experiment, our method outperforms state-of-the-art models in both high-resource and low-resource settings.

Hazes with various properties spread widely across flat areas with depth continuities and corner areas with depth discontinuities. Removing haze from a single hazy image is difficult due to its ill-posed nature. To solve this problem, this study proposes a modified hybrid median filter that performs a median filter to preserve the edges of flat areas and a hybrid median filter to preserve depth discontinuity corners. Recovered scene radiance, which is obtained by removing hazy particles, restores image visibility using adaptive nonlinear curves for dynamic range expansion. Using comparative studies and quantitative evaluations, this study shows that the proposed method achieves similar or better results than those of other state-of-the-art methods.

We propose a heuristic approximation algorithm for the 1-median problem. The 1-median problem is the problem of finding a vertex with the highest closeness centrality. Starting from a randomly selected vertex, our algorithm repeats to find a vertex with higher closeness centrality by approximately calculating closeness centrality of each vertex using simpler spanning subgraphs, which are called k-neighbor dense shortest path graphs with shortcuts. According to our experimental results using real networks with more than 10,000 vertices, our algorithm is more than 100 times faster than the exhaustive search and more than 20 times faster than the state-of-the-art approximation algorithm using annotated information to the vertices while the solutions output by our algorithm have higher approximation ratio.

In this letter, we propose a simple framework for accelerating a state-of-the-art histogram-based weighted median filter at no expense. It is based on a process of determining the filter processing direction. The determination is achieved by measuring the local feature variation of input images. Through experiments with natural images, it is verified that, depending on input images, the filtering speed can be substantially increased by changing the filtering direction.

A new technology for video frame rate up-conversion (FRUC) is presented by combining a median filter and motion estimation (ME) with an occlusion detection (OD) method. First, ME is performed to obtain a motion vector. Then, the OD method is used to refine the MV in the occlusion region. When occlusion occurs, median filtering is applied. Otherwise, bidirectional motion compensated interpolation (BDMC) is applied to create the interpolated frames. The experimental results show that the proposed algorithm provides better performance than the conventional approach. The average gain in the PSNR (Peak Signal to Noise Ratio) is always better than the other methods in the Full HD test sequences.

This paper proposes a robust bilateral filter which can handle mixed Gaussian and impulsive noise by hybridizing the conventional bilateral filter and the switching median filter. The effectiveness of the proposed method is verified in comparison with other conventional methods by some experiments using the natural digital images.

In this paper, we propose a robust registration method, named Bounded-Variables Least Median of Squares (BVLMS). It overcomes both the misassociations and the ill-conditioning due to the interactions between Bounded-Variables Least Squares (BVLS) and Least Median of Squares (LMS). Simulation results demonstrate the feasibility of this new registration method.

In this paper, we propose a new motion vector smoothing algorithm using weighted vector median filtering based on edge direction for frame interpolation. The proposed WVM (Weighted Vector Median) system adjusts the weighting values based on edge direction, which is derived from spatial coherence between the edge direction continuity of a moving object and motion vector (MV) reliability. The edge based weighting scheme removes the effect of outliers and irregular MVs from the MV smoothing process. Simulation results show that the proposed algorithm can correct wrong motion vectors and thus improve both the subjective and objective visual quality compared with conventional methods.

In the latest video coding frameworks, efficiency of motion vector (MV) coding is becoming increasingly important because of the growing bit rate portion of motion information. However, neither the conventional median predictor, nor the newer schemes such as the minimum bit rate prediction scheme and the hybrid scheme, can effectively eliminate the local redundancy of motion vectors. In this paper, we present the prioritized reference decision scheme for efficient motion vector coding, based on the H.264/AVC framework. This scheme makes use of a boolean indicator to specify whether the median predictor is to be used for the current MV or not. If not, the median prediction is considered not suitable for the current MV, and this information is used for refining the possible space of a group of reference MVs including 4 neighboring MVs and the zero MV. This group of MVs is organized to be a prioritized list so that the reference MV with highest priority is to be selected as the prediction value. Furthermore, the boolean indicators are coded into the modified code words of mb_type and sub_mb_type, so as to reduce the overhead. By applying the proposed scheme, the structure and the applicability problems with the state-of-the-art MBP scheme have been overcome. Experimental result shows that the proposed scheme achieves a considerable reduction of bits for MVDs, compared with the conventional median prediction algorithm. It also achieves a better and much stabler performance than MBP-based MV coding.

For the acquisition of visual information, the nonuniform sampling process by photoreceptors on the retina occurs at the earliest stage of visual processing. From objects of interest, the human eye receives high visual resolution through nonuniform distribution of photoreceptors. Therefore, this paper proposes auto exposure and focus algorithms for the real-time video camera system based on the visual characteristic of the human eye. For given moving objects, the visual weight is modeled for quantifying the visual importance and the associated auto exposure and focus parameters are derived by applying the weight to the traditional numerical expression, i.e., the DoM (Difference of Median) and Tenengrad methods for auto focus.

Auto-calibration for structure and motion recovery can be used for match move where the goal is to insert synthetic 3D objects into real scenes and create views as if they were part of the real scene. However, most auto-calibration methods for multi-views utilize bundle adjustment with non-linear optimization, which requires a very good starting approximation. We propose a novel key-frame selection measurement and LMedS (Least Median of Square)-based approach to estimate scene structure and motion from image sequences captured with a hand-held camera. First, we select key-frames considering the ratio of number of correspondences and feature points, the homography error and the distribution of corresponding points in the image. Then, by using LMedS, we reject erroneous frames among the key-frames in absolute quadric estimation. Simulation results demonstrated that the proposed method can select suitable key-frames efficiently and achieve more precise camera pose estimation without non-linear optimization.

In this paper, we present a directional interpolation filter in which the minimum and maximum pixels in the given window are excluded. Image pixels within a predefined window are ranked and classified as minimum-maximum or exclusive level, and then passed through the interpolation and identity filters, respectively. Extensive simulations show that the proposed filter performs better than other nonlinear filters in preserving desired image features while reducing impulse noise effectively.

A new impulse noise detection algorithm is presented, which can successfully remove impulse noise from corrupted images while preserving image details. The impulse detection algorithm is combined with median filtering to achieve noise removal. The main advantage of the proposed algorithm is that it can detect the impulse noise with high accuracy while reducing the probability of detecting image details as impulses. Also, it can be applied iteratively to improve the quality of restored images. It is efficient and low in complexity. Furthermore, it requires no previous training. Extensive experimental results show that the proposed approach significantly outperforms many well-known techniques.

In this paper, we present a minimum-maximum exclusive weighted-mean filtering algorithm with adaptive window. Image pixels within the varying size of the window are ranked and classified as minimum-maximum and median levels, and then passed through the weighted-mean of median level and identity filters, respectively. The filtering window size is adaptively increasing according to noise ratio without noise measurement. Extensive simulations show that the proposed filter performs better than other median/rank-type filters in removing impulse noise of highly corrupted images.

A new sorting algorithm and architecture for fast median filter are proposed. This algorithm results in low area VLSI architecture producing low switching activity and without using feedback. The main idea is to employ the extra matrix for fast search operation of rank of oldest window element. We simulated and synthesized this algorithm using SYNOPSYSTM and showed the sufficiency in real time operation.

A novel effective method for detection and removal impulse noise in highly corrupted color images is proposed. This detection-estimation method consists of two steps. Outliers are first detected using spatial relations between the color components. Then the detected noise pixels are replaced with the output of the vector median filter over a local spatially connected area excluding the outliers. Simulation results in a test color image show a superior performance of the proposed filtering algorithm comparing to the conventional vector median filter. The comparisons are made using a mean square error and a mean absolute error criteria.

In this study, we consider a filtering method for image sequence degraded by additive Gaussian noise and/or impulse noise (i.e., mixed noise). For removing the mixed noise from the 1D/2D signal, weighted median filters are well known as a proper choice. We have also proposed a filtering tool based on the weighted median filter with a data-dependent method. We call this data-dependent weighted median (DDWM) filters. Nevertheless, the DDWM filter, its weights are controlled by some local information, is not enough performance to restore the image sequence degraded by the noise. The reason is that the DDWM filter is not able to obtain good filtering performance both in the still and moving regions of an image sequence. To overcome above drawback, we add motion information as a motion detector to the local information that controls the weights of the filters. This new filter is proposed as a Video-Data Dependent Weighted Median (Video-DDWM) filter. Through some simulations, the Video-DDWM filter is shown to give effective restoration results than that given by the DDWM filtering and the conventional filtering method with a motion-conpensation (MC).

In this paper, we propose a fast algorithm to realize parallel median filter for processing 1-D and 2-D signal. In the proposed pipelined architecture, m-passes are employed for filtering signal while word resolution is m bits. One pass employs one processing element (PE), and the number of PEs is independent of the number of samples. Therefore, we only need m PEs for real-time operation. With 8-bits resolution, the system gate-count is less than 5 k. Moreover, this median architecture could be easily modified to consist of the programmable feature that may choose the better sampling number to filter signal. It should be also noted that our proposed processing flow has a progressive property, which is very suitable for bandwidth-limited channel application.

In this paper, an algorithm of the median-cut quantization (MCQ) is proposed. MCQ is the technique that reduces multi-valued samples to binary-valued ones by adaptively taking the median value as the threshold. In this work, the search process of the median value is derived from the quick-sort algorithm. The proposed algorithm searches the median value bit by bit, and samples are quantized during the search process. Firstly, the bit-serial procedure is shown, and then it is modified to the bit-parallel procedure. The extension to the multi-level quantization is also discussed. Since the proposed algorithm is based on bit operations, it is suitable for hardware implementation. Thus, its hardware architecture is also proposed. To verify the significance, for the application to the motion estimation, the performance is estimated from the synthesis result of the VHDL model.

Visualization of 3-D ultrasound images is a challenging task due to the noisy and fuzzy nature of ultrasound imaging. This paper presents an efficient volume rendering technique for 3-D ultrasound image. A preprocessing technique of 2-D truncated-median filtering is proposed to reduce speckle noise of the ultrasound image. This paper also introduces an adaptive boundary detection method to reduce the computation time for volume rendering of ultrasound image. The proposed technique is compared to the conventional volume rendering methods with respect to the computation time and the subjective image quality. According to the comparison study, the proposed volume rendering method shows good performance for visualization of 3-D ultrasound image.