Search

Glaucoma is a neurodegenerative disease presents with retinal nerve fiber layer (RNFL) defects. We apply machine learning classifiers on the color information of the RNFL to differentiate between intact RNFL (i-RNFL) and RNFL defect (d-RNFL) on optic disc photographs (DPs). DPs from individuals with and without glaucoma were collected. Then, a semi-circle was automatically marked on the DPs, to label i-RNFL versus d-RNFL. RGB intensities and other color spaces of two profiles were collected. Five-fold cross validation is used to compare classification efficiency of five classifiers. A total of 2,051 profiles from 89, 32 and 15 DPs from patients with glaucoma, glaucoma suspects and control subjects were collected. There were 702 and 175 points of d-RNFL and 940 and 234 of i-RNFL in the training and test sets. In the training set, the 3 best classifiers using RGB intensities were fine Gaussian support vector machine (SVM), medium k-Nearest Neighbor and ensemble RUSBoosted Trees, with accuracies of 81.8%, 79.4% and 79.2%. The performance of the fine Gaussian SVM was similar between RGB and other color spaces. In the test set, the highest sensitivity (71.4%) and specificity (88.5%) were archived using RGB and the combination of RGB and Cb and Cr.

Broiler is among Thai major industries. Red blood cell (RBC) count is one of the basic health screening protocols. The counting is often done manually. The mammalian blood analyzer cannot be used, because the shapes of mammalia and avian RBCs are different. RBC's color also varies among staining methods. Furthermore, RBCs can be grouped together. In this paper, we propose the automatic method for counting avian RBC. Otsu's multiple thresholding is used to extract the nucleus of RBC. The problems of the varying color of and the conglomerated RBC are tackled by applying iterative process. Only one parameter needs to be set manually. The experiment on five different staining types demonstrated that our proposed method provided less than 5% counting error for the staining that did not have the dark RBC's boundary.

The chicken industry ranks tenth in the world in term of output in the world meat market. Red blood cell (RBC) count is one of the basic health screening protocol required in an exported meat industry. However, the human’s automated blood analyzer cannot be applied to avian RBC, because the shape of avian blood (ellipse) is different from the mammal one (circle). In this paper, we propose 2-stage thresholding technique for automatic avian RBC counting. First, the blood smear slide is binarized into RBC and non-RBC area by applying Otsu thresholding. Then, image morphology and Otsu thresholding are reapplied to detect the blood nucleus. After that, the connected component analysis is applied to count the number of RBC. The experiment demonstrated that the proposed technique was simple and provided the count with error rate (2.23%) much less than the clinically acceptable value (5%).

Red blood cell counting is one of the most requested blood tests and can be effectively done by the dedicated equipment. In this paper, we proposed the two-stage circular Hough transform as the alternative automatic method for canine red blood cell counting. The possible blood cell area in the input slide image is first extracted. In the first stage, the Circular Hough transform with low sensitivity is applied to detect the non-overlapped cells. The detected cells are removed. The Circular Hough transform with high sensitivity is re-applied to the remaining region in order to detect the overlapped cells. The proposed method can be easily implemented into personal computers. The proposed algorithm was tested with many slides of canine red blood cell. It provided the counting with the average error of less than 5% with the best result of 1.439% error.

One of the fundamental noise in Computed Tomography (CT) imaging is metal artifact. The artifact leads to the information loss. Magnetic Resonance (MR) imaging is not affected by this artifact. In this paper, we propose using the information from MR images to reproduces the information lost due to dental filling in CT images. First, we register the CT and MR images of the same patient. Then, we classify the skull into 3 regions: teeth, tissue near the teeth and others. For teeth and tissue near the teeth, we develop the mapping function between the intensity in the MR image to the mean intensity in the CT image. The pixels corrupted by the metal artifact are detected and their intensity are changed according to the mapping function with its MR intensity as the input. The proposed method significantly removed metal artifacts and could be used to improve the diagnosis and planning accuracy.

Directional Local Mean Difference Level Set method (DLMD-LS) is the segmentation method for a urinary bladder in an MR sequence used for planning the treatment of a cervical cancer by radiation. The blurred boundary of a bladder is segmented based on the judgment of a radiologist and can be differed among radiologists. In this paper, DLMD-LS with Reinforcement Learning (RL) is proposed. It is an interactive system, where the parameter is adjusted to reflect the individual judgment. The weighted average method is used to update the distance that the boundary will be expanded, after the level set contour finishes evolving. The experiment on 30 MR slices demonstrated that DLMD-LS with RL had high segmentation accuracy and was adaptable to the new radiologist. It was also robust to outliers.

In this paper, we propose the directional local mean difference level set method (DLMD-LS). Our work is focused on the segmentation for a urinary bladder lumen in a T2 weighted image taken during brachytherapy. The boundary is detected as the region where the intensity means of the areas inside and outside the zero-level contour (edge) are high. The contour is controlled such that it stops evolving only at the boundary of the bright object inside the dark surrounding. Since the intensity mean is used, the DLMD-LS is more tolerant to noise than the conventional level set methods which detect the boundary as the part with large intensity gradient. Furthermore, the mean is calculated locally; therefore, the function is less affected to the different intensity distribution inside the lumen. The experiment demonstrates the superior noise tolerance and better capability to capture the blurred boundary of our proposed DLMD-LS to the directional level set method and the level set which detects the boundary according to the intensity distribution of the entire lumen.

Impulsive noise in a compressed measurement signal can be detected by the energy distribution of the reconstructed sparse signal. In this paper, the fast preprocessing for impulsive noise rejection for a compressed measurement image signal is proposed. An image is sparsified by octave-tree wavelet transform with db8 as the mother wavelet. The impulsive noise is detected via the change in the ratio of the energy leaking out of the third level subband to the total energy of the sparse signal. Since the information in the image is redundant, the number of the noisy element is estimated within +g of the actual value. g is defined in the unit of the percent of the size of the compressed measurement signal. The experiment shows that the proposed method effectively rejected noise with large magnitude. It provided the comparable performance to the method that rejected the exact number of noisy elements but required less computational time.

This paper proposes a formulation of estimating a sparse image signal where the sparsity only occurs on the high frequency components. The formulation is for compressed sensing reconstruction when a compressed measurement signal is contaminated by impulsive noise. The approach is based on two prior well-known assumptions. First, adding an L 1 -norm penalty on a compressible signal to the cost objective of the estimation problem will promote a sparsity in the signal. Secondly, an image signal should be sparse in high frequency domain, while it is dense in low frequency domain where most of its important information lies. Our approach is therefore to impose an L 1 -norm penalty only on the signal components that need to be sparse while we allow the important low frequency part to be dense. The obtained formulation is a convex optimization which can be effectively solved by many available algorithms. The experiment demonstrated a superior impulsive noise tolerance of the proposed penalty function to a conventional scheme where all signal components are penalized by L 1 -norm. At the noise probability of less than 0.1, the reconstruction was better both quantitatively and qualitatively.

Impulsive noise in a compressed measurement signal, y, has a significant effect on the reconstruction in Compressed Sensing (CS). In this paper, Greedy gap's Boundary Finder (GBF), a fast preprocessing for impulsive noise rejection for the CS reconstruction of image signal, is proposed. An image is sparsified by the octave-tree wavelet transform. GBF adopts an idea that the leakage energy out of the third-level (L 3 ) subband in the reconstructed sparse signal is the result of either the impulsive noise in y or the insufficient information (of y) for reconstruction. The leakage energy is measured as the ratio to the total energy. In a graph of relationship between the energy ratio and the number of the removed elements, there is a gap of low energy ratio in the middle. A binary search is adapted in GBF to find the lower gap's boundary which is the minimum number of removed elements with low energy ratio. As the image signal is highly redundant, the search in GBF can be non-exact and will be stopped after the number is within +g of the lower boundary, where g is the gap's resolution and defined as the percent of the size of y. GBF was evaluated on 100 16×16 image blocks and 10 256×256 standard test images. The evaluation shows that at the optimal g of 5%, GBF provided the comparable performance to the exact boundary's search, but consumed less computational time.

A reconstruction method using the ensemble of compressed measurement signals is proposed for reconstructing the image from the signal corrupted by Gaussian noise. The ensemble is created from one signal under the assumption that an image is highly redundant; hence, it is approximated as the mixture of a number of signals. The proposed method adopted the sampling with replacement in bootstrapping to extract L signals from the mixture. The extracted L signals from the ensemble of signals corrupted by Gaussian noise with the same mean and variance. The signals have different bases; thus, they he in different space. Orthogonal Matching Pursuit with Partially Known Support (OMP-PKS) is applied to reconstruct the L signals to the same sparse space. Gaussian noise is reduced by averaging the reconstructed signals. The performance of the proposed method was compared with Basis Pursuit Denoising (BPDN), OMP-PKS and Distributed Compressed Sensing using Simultaneously Orthogonal Matching Pursuit (DCS-SOMP). The experimental results of 10 standard test images showed that our method yielded higher Peak Signal-to-Noise Ratio (PSNR) and better visual quality at a high level of noise.

In image stitching, parallax is key challenge that leads to inaccurate registration and ghosting effect of objects in the result. A novel foreground rejection method is proposed in this paper to remove parallax in video sequence. Firstly, the global motion is estimated between two frames using SIFT feature matching. The logical OR operators are then applied to the pixels that have high displaced frame difference for foreground extraction. There are two sets of foreground calculated from the near frames and far frames. Voting scheme is applied in the way that only the near-frame foreground appears inside the area of far-frame foreground that is considered as the actual foreground. The extracted foreground at this stage is mostly the edges of objects. Therefore, the foreground rejection method proposed in this paper aims to provide better motion vector estimation and ghosting effect reduction. The removal of the major parts of the high difference foreground is sufficient; hence the simple two-stage rendering is applied. The implemented results indicate that more accurate registration is achieved by including our proposed foreground rejection method. Moreover, the background of the stitched video sequence does not have the visible ghosting effect.

In video frame rate conversion, unavailable frames are reconstructed from existing frames based on interpolated motion vectors. Due to an error in estimating motion vectors and an occlusion effect, it is possible that intensity information is not provided to every pixel in the reconstructed frame. Pixels whose intensity is missing become holes in the reconstructed frame. This paper proposes the reconstruction without distinct holes by double motion vector estimation. Motion is described by affine model instead of the conventional 2D translation model. The proposed method applies block based motion estimation. It is divided into 2 stages: frame reconstruction and hole filling. In the reconstruction stage, a new frame is reconstructed by moving blocks in previous frames according to their interpolated motion vectors. Holes in the newly reconstructed frame are then filled with the corresponding pixels in the subsequent frame. The proposed method searches for the corresponding pixels by looking for the region in the subsequent frames that are similar to the surrounding region of each hole. The computation cost in the proposed method is low because the motion estimation is simple and applied only twice. It does not use any complex processes to fill holes. Experiments on three standard video sequences showed that our proposed method reconstructed frames in complex video sequence without noticeable holes.

Due to large data set, block processing is usually applied for fast compressive sensing (CS) reconstruction; however, it gives the undesired blocking artifact in reconstructed data. In order to reduce blocking artifact and preserve high frequency, this paper proposes a novel block processing on wavelet domain instead of spatial domain. No post-processing nor special mapping is included. CS is applied only on blocks where the data are really sparse. An enhancing process, often included for artifact reduction, is no longer necessary. Our algorithm was evaluated by reconstructing three standard images (Lena, Mandrill and Peppers) and then compared with scrambled block hadamard ensemble (SBHE) and block-based CS sampling with a smoothed PL variant using directional discrete wavelet transform (BCS-SPL-DDWT). In the experiment, it provided better reconstruction both objectively (PSNR) and subjectively at low measurement rate. It gave the sharpest image in all cases. Details were preserved and blocking artifact was not detectable.

Traditionally, the problems of applying Orthogonal Matching Pursuit (OMP) to large images are its high computing time and its requirement for a large matrix. In this paper, we propose a fast image recovery algorithm by dividing the image into block of n×n pixels and applying OMP to each n×n block instead of the entire image. The key idea is that small matrix requires less computing time and less memory. In the experiment, the block based OMP was applied to three standard test images: Lena, Mandrill and Pirate. Compared to standard OMP, block based OMP required less computing time while giving comparable PSNR.

The quality of the frame rate up-conversion depends on the accuracy of the motion parameters. 2D affine model describes the motion better than the popular 2D translation model; however, its parameters are difficult to interpolate. In this paper, the approximation method for interpolating the mid-point of 2D affine parameters is proposed. The affine transform is approximated as the 2D rigid transformation where the target x and y axes are not orthogonal. Trigonometry relationship is applied to interpolate rotational parameters. Conventional linear interpolation is applied to interpolate translational parameters. For video sequences containing complex motion, the reconstruction from the approximated affine parameters gave higher PSNR and visually better result than the one from the interpolated 2D translation parameters.

Most of segmentation algorithms are based on the assumption of intensity homogeneity within an object. However, in many applications, the object of interest contains more than one homogenous-intensity region. Even when the object's shape is known, such object is not effectively extracted. In this paper, we propose a segmentation process for the objects containing 2 homogenous regions. Our method is based on level set method. This method is divided into 2 parts. In the first part, we estimate the pose parameters of the mean shape of the object. In the latter part, the region inside the curve is adjusted to form a homogenous region. Then we segment the adjusted image for the object by level set method. Our experiment on medical images indicated that our method effectively segmented object with one and two homogenous regions.

Most segmentation algorithms are based on the assumption of intensity homogeneity within an object. However, in many applications, the object of interest contains more than one homogenous region. Even when the object’s shape is known, such object is not effectively extracted. In this paper, we propose a segmentation process for the objects containing 2 homogenous regions. Our method is based on the level set method. We construct the shape model from the set of manually extracted objects. The parameters that represent the shape model are coefficients of PCA basis. Instead of defining a new cost-function based on heterogeneity assumption, we repeatedly form a homogenous region inside the evolving curve and evolve the curve by the level set method. Our experiment on medical images indicated that our method effectively segmented object with one and two homogenous regions.