Large Filter Kernel Research Articles

Automated Detection of Age-Related Macular Degeneration from OCT Images Using Multipath CNN

Age-related macular degeneration (AMD) is an eye disorder that can have harmful effects on older people. AMD affects the macula, which is the core portion of the retina. Hence, early diagnosis is necessary to prevent vision loss in the elderly. To this end, this paper proposes a novel multipath convolutional neural network (CNN) architecture for the accurate diagnosis of AMD. The architecture proposed is a multipath CNN with five convolutional layers used to classify AMD or normal images. The multipath convolution layer enables many global structures to be generated with a large filter kernel. In this proposed network, the sigmoid function is used as the classifier. The proposed CNN network is trained on the Mendeley dataset and evaluated on four datasets-the Mendeley, OCTID, Duke, and SD-OCT Noor datasets- and it achieved accuracies of 99.60%, 99.61%, 96.67%, and 93.87%, respectively. Although the proposed model is only trained on the Mendeley dataset, it achieves good detection accuracy when evaluated with other datasets. This indicates that the proposed model has the capacity to detect AMD. These results demonstrate the efficiency of the proposed algorithm in detecting AMD compared to other approaches. The proposed CNN can be applied in real-time due to its reduced complexity and learnable parameters.

ОЦІНКА ЕФЕКТИВНОСТІ МЕТОДІВ САМООЦІНЮВАННЯ ЗОБРАЖЕННЯ

Lighting Normalization is an especially important issue in the image recognitions systems since different illumination conditions can significantly change the recognition results, and the lighting normalization allows minimizing negative effects of various illumination conditions. In this paper, we are evaluating the recognition performance of several lighting normalization methods based on the Self-Quotient ImagE(SQI) method introduced by Haitao Wang, Stan Z. Li, Yangsheng Wang, and Jianjun Zhang. For evaluation, we chose the original implementation and the most perspective latest modifications of the original SQI method, including the Gabor Quotient ImagE(GQI) method introduced by Sanun Srisuk and Amnart Petpon in 2008, and the Fast Self-Quotient ImagE(FSQI) method and its modifications proposed by authors in previous works. We are proposing an evaluation framework which uses the Cropped Extended Yale Face Database B, which allows showing the difference of the recognition results for different illumination conditions. Also, we are testing all results using two classifiers: Nearest Neighbor Classifier and Linear Support Vector Classifier. This approach allows us not only to calculate recognition accuracy for each method and select the best method but also show the importance of the proper choice of the classification method, which can have a significant influence on recognition results. We were able to show the significant decreasing of recognition accuracy for un-processed (RAW) images with increasing the angle between the lighting source and the normal to the object. From the other side, our experiments had shown the almost uniform distribution of the recognition accuracy for images processed by lighting normalization methods based on the SQI method. Another showed but expected result represented in this paper is the increasing of the recognition accuracy with the increasing of the filter kernel size. However, the large filter kernel sizes are much more computationally expensive and can produce negative effects on output images. Also, we were shown in our experiments, that the second modification of the FSQI method, called FSQI3, is better almost in all cases for all filter kernel sizes, especially, if we use Linear Support Vector Classifier for classification.

Deep Color Guided Coarse-to-Fine Convolutional Network Cascade for Depth Image Super-Resolution.

Depth image super-resolution is a significant yet challenging task. In this paper, we introduce a novel deep color guided coarse-to-fine convolutional neural network (CNN) framework to address this problem. First, we present a datadriven filter method to approximate the ideal filter for depth image super-resolution instead of hand-designed filters. Based on large data samples, the filter learned is more accurate and stable for upsampling depth image. Second, we introduce a coarse-to-fine CNN to learn different sizes of filter kernels. In coarse stage, larger filter kernels are learned by CNN to achieve crude high-resolution depth image. As to fine stage, the crude high-resolution depth image is used as the input so that smaller filter kernels are learned to gain more accurate results. Benefit from this network, we can progressively recover the high frequency details. Third, we construct a color guidance strategy that fuses color difference and spatial distance for depth image upsampling. We revise the interpolated high-resolution depth image according to the corresponding pixels in highresolution color maps. Guided by color information, the depth of high-resolution image obtained can alleviate texture copying artifacts and preserve edge details effectively. Quantitative and qualitative experimental results demonstrate our state-of-the-art performance for depth map super-resolution.

Compact precomputed voxelized shadows

Producing high-quality shadows in large environments is an important and challenging problem for real-time applications such as games. We propose a novel data structure for precomputed shadows, which enables high-quality filtered shadows to be reconstructed for any point in the scene. We convert a high-resolution shadow map to a sparse voxel octree, where each node encodes light visibility for the corresponding voxel, and compress this tree by merging common subtrees. The resulting data structure can be many orders of magnitude smaller than the corresponding shadow map. We also show that it can be efficiently evaluated in real time with large filter kernels.

Wavefront raycasting using larger filter kernels for on-the-fly GPU gradient reconstruction

The quality of images generated by volume rendering strongly depends on the accuracy of gradient estimation. However, the most commonly used techniques for on-the-fly gradient reconstruction are still very simple, such as central differences; they generally gather only limited neighbourhood information and thus ultimately produce rather poor quality images. While there are many higher-order reconstruction methods, such as 3×3×3 or 5×5×5 filters, which can improve the quality, their excessive sampling costs have meant that they are generally used only for pre-computed gradients, which are then quantized and stored for later runtime re-interpolation. This may introduce further errors and, significantly, may consume valuable texture memory. In this paper, we address these issues by proposing a CUDA-based rendering framework that uses larger filter kernels for on-the-fly gradient computation in real-time raycasting applications. By using adaptive wavefront tracing, our approach can dramatically reduce the memory bandwidth requirements related to larger neighbour samples. To further ensure that samples are consumed wisely, we have devised a novel adaptive sampling scheme and a customized 3D mipmapping technique in the CUDA environment to sample at a proper level of detail as the ray recedes into the distance. We compared our technique with two previous state-of-the-art GPU raycasting algorithms and found that it achieves higher quality imaging and faster rendering performance across a variety of data sets than the previous methods.

Effects of natural backgrounds on spatial filter responses near object contours

Purpose: The description of an object's contour is often thought to be solely a property of the object's image. However, objects typically appear in complex backgrounds. Thus for any 2D sample region (SR), both the preferred orientation and the magnitude of spatial contrast filters can be significantly perturbed by the background. The degree of perturbation is a function of filter size and the position of the filter relative to the contour. Further, the response of larger filters may be less sensitive to positional uncertainty. Our goal was to measure this perturbation by natural backgrounds. Methods: Images of natural and artificial objects were photographed and hand-segmented from their natural backgrounds. Composite images were made by digitally superimposing and translating the image of an object across various natural image backgrounds. 2D histograms of orientation and magnitude responses were measured from a set of steerable multi-scale filters for SRs across contours and on the background. In order to quantify discriminability, we measured the performance of an ideal observer. The ideal was required to discriminate between specific contour and background SRs, given prior knowledge of the conditional distributions estimated from the histograms. Results: Histograms of SRs on contours varied systematically as a function of displacement from the contour and were distinctly different than those from the background. The performance of the ideal observer was surprisingly good, with an overall correct response rate of .78. When centered on the contour, correct response rate was .82. Larger filter kernels were less sensitive to centering on the contour. Conclusions: Background has a significant effect on contour filter responses. However, these perturbations form predictable patterns, and allow discrimination with reasonable accuracy. Filter size helps to minimize spatial uncertainty.

Sum-box technique for fast linear filtering

Linear filters are widely used in signal processing and in image processing, and the fast realization of large kernel filters has been an important research subject. Traditional fast algorithms try to decompose the initial kernel into the convolution of smaller kernels, and a lot of multiplications are still needed to realize all the convolutions with the small kernels. Box technique can realize the Gaussian filters with little multiplications, but it cannot realize linear filters other than the Gaussians. In the present paper, by use of the analysis on the scaled Spline functions, we propose the sum-box filter technique to approximately realize a given large kernel linear filter (even non Gaussian type) to a factor by the sum of the translated outputs of sum-box filters, requiring no multiplications. This original sum-box technique presented opens thus a new orientation for linear filter fast realization: realizing to a factor the convolution with a large filter kernel by additions only. From the viewpoint of computational complexity, this method is highly efficient in image and signal processing, in particular, for large filter kernels. Experimental results are reported.

Frequency-filtered image post-processing of digital luminescence radiographs in pulmonary nodule imaging

Aim: The aim of the study was to optimize unsharp masking image post-processing of digital luminescence radiographs (DLR) for the representation of pulmonary nodules, and to compare DLR to screen-film radiography at two dose levels. Patients and methods: A total of 284 CT-validated pulmonary nodules were evaluated. One hundred and forty-nine nodules were exposed with a 200-speed screen-film combination (SFC) and 135 nodules with a 400-speed SFC, with correspondingly exposed storage phosphor images. The kernal size in digital post-processing using ‘unsharp masking’ was varied between S 10 (2.83mm) and S 70 (19.80mm). A total of 11 928 individual assessments were obtained from six independent observers and evaluated in multifactorial variance analyses. Results: The large filter kernels of S 40 and S 70 were on a par with the 200-speed SFC ( P > 0.05). As the exposure dose was reduced, the quality of the digital image vis-à-vis the 400-speed SFC improved significantly ( P < 0.05). Smaller filter kernels (S 10; S 20) producing edge-enhancement processing were significantly inferior to the analog image technique in both dose ranges ( P < 0.05). Conclusions: At speed class 200, low-frequency emphasizing digital image post-processing with large filter kernels are significantly superior to high-frequency emphasizing filtrations for the recognition of pulmonary nodules. In the lower dose range DLR with large filter kernel unsharp masking processing showed significantly improved image quality compared to 400 speed SFC for the detection of pulmonary nodules.

ROC-Analyse zur Bildnachverarbeitung digitaler Thoraxaufnahmen

To examine the extent to which digital luminescence radiography (DLR) can be used for the imaging of pulmonary nodules and interstitial lung disease in chest radiography without any loss of image quality. Additionally: to examine whether post-processing of image data can optimise the recognizability of varied image details. Detail perceptibility studies were performed on an anthropomorphic thorax phantom with simulated nodules and small linear and reticular details. Under standard conditions, digital luminescence radiographs were obtained in 7 different image modes, and these were compared with a 200-speed screen-film system. The detection of these systems was evaluated in an ROC analysis on the basis of 19,200 individual observations. Edge enhancement or application of high-frequency-enhancing small filter kernels (S 5) slightly improves the detection of linear structures; however, the illustration of nodular details is markedly reduced. Larger filter kernels (S 20, S 40) make a definitive detection possible--not only of circular, but also of linear details. Storage phosphor radiographs are equal to the tested analog screen-film-system. The optimization of post-processing can be helpful in the prevention of routine multiple documentations.