Pattern Analysis and Applications

Training deep neural networks (DNNs) on high-dimensional data with no spatial structure poses a major computational problem. It implies a network architecture with a huge input layer, which greatly increases the number of weights, often making the training infeasible. One solution to this problem is to reduce the dimensionality of the input space to a manageable size, and then train a deep network on a representation with fewer dimensions. Here, we focus on performing the dimensionality reduction step by randomly projecting the input data into a lower-dimensional space. Conceptually, this is equivalent to adding a random projection (RP) layer in front of the network. We study two variants of RP layers: one where the weights are fixed, and one where they are fine-tuned during network training. We evaluate the performance of DNNs with input layers constructed using several recently proposed RP schemes. These include: Gaussian, Achlioptas’, Li’s, subsampled randomized Hadamard transform (SRHT) and Count Sketch-based constructions. Our results demonstrate that DNNs with RP layer achieve competitive performance on high-dimensional real-world datasets. In particular, we show that SRHT and Count Sketch-based projections provide the best balance between the projection time and the network performance.

Adaptive localizing region-based active contour model driven by Laplacian kernel-based fitting energy is proposed for improving the efficiency and accuracy of synthetic aperture radar (SAR) river image segmentation in the paper. Defining regional energy functional that depends on the Laplacian kernel distance which is robust and non-Euclidean, Laplacian kernel distance is nonlinear transformation, whose transformed space can be linear classification. Additionally, providing the novel calculation for fitting center which relies on the local and global gray value, furthermore, the adaptive selection function of local radius is made. By using both of them, the proposed model can improve the accuracy of the fitting center and local region; afterward, the evolution of the curve can achieve the global optimal and be controlled better. Finally, in order to speed up the computation of proposed model, the localized region surrounded by adjacent four pixel points on the evolution curve can be replaced by the localized region of the intermediate pixel. The proposed model has been successfully applied to river channel extraction from synthetic aperture radar (SAR) images with desirable results. Comparisons with other state-of-the-art approaches demonstrate the great performances of the model.

In the last decade, computer vision, pattern recognition, image processing and cardiac researchers have given immense attention to cardiac image analysis and modelling. This paper survets state-of-the-art computer vision and pattern recognition techniques for Left Ventricle (LV) segmentation and modelling during the second half of the twentieth century. The paper presents the key charateristics of successful model-based segmentation and modelling during the second half of the twentieth century. The paper presents the key characteristics of successful model-based segmentation techniques for LV modelling. This survey paper concludes the following: (1) any one pattern recognition or computer vision technique is not sufficient for accurate 2D, 3D or 4D modelling of LV; (2) fitting mathematical models for LV modelling have dominated in the last 15 tears; (3) knowledge extrated from the ground truth has lead to very successful attempts for LV modelling have dominated in the last 15 uears; (3) knowledge extracted from the ground truth has lead to very successful attempts at LV modelling;(4) spatial and temporal behaviour of LV through different imaging modalities has yielded information which has led to accurate LV modelling; and (5) not much attention has been paid to LC modelling validation.

Magnetic resonance (MR) tomographic images are routinely used in diagnosis of liver pathologies. Liver segmentation is needed for these types of images. It is therefore an important requirement for later tasks such as comparison among studies of different patients, as well as studies of the same patient (including those taken during the diffusion of a contrast, as in perfusion MR imaging). However, automatic segmentation of the liver is a challenging task due to certain reasons such as the high variability of liver shapes, similar intensity values and unclear contours between the liver and surrounding organs, especially in perfusion MR images. In order to overcome these limitations, this work proposes the use of a probabilistic atlas for liver segmentation in perfusion MR images, and the combination of the information gathered with that provided by level-based segmentation methods. The process starts with an under-segmented shape that grows slice by slice using morphological techniques (namely, viscous reconstruction); the result of the closest segmented slice and the probabilistic information provided by the atlas. Experiments with a collection of manually segmented liver images are provided, including numerical evaluation using widely accepted metrics for shape comparison.

This paper presents a real-time vision-based system to detect the eye state. The system is implemented with a consumer-grade computer and an uncalibrated web camera with passive illumination. Previously established similarity measures between image regions, feature selection algorithms, and classifiers have been applied to achieve vision-based eye state detection without introducing a new methodology. From many different extracted data of 1,293 pair of eyes images and 2,322 individual eye images, such as histograms, projections, and contours, 186 similarity measures with three eye templates were computed. Two feature selection algorithms, the $$ J_{5} (\xi ) $$ criterion and sequential forward selection, and two classifiers, multi-layer perceptron and support vector machine, were intensively studied to select the best scheme for pair of eyes and individual eye state detection. The output of both the selected classifiers was combined to optimize eye state monitoring in video sequences. We tested the system with videos with different users, environments, and illumination. It achieved an overall accuracy of 96.22 %, which outperforms previously published approaches. The system runs at 40 fps and can be used to monitor driver alertness robustly.

In this paper, an algorithm for robust finger vessel pattern extraction from infrared images is presented based on image processing, edge suppression, fuzzy enhancement, and fuzzy clustering. Initially, the brightness variations of the images are eliminated using histogram normalization and the vessel patterns are enhanced, to facilitate the separation process from other tissue parts through fuzzy clustering. Several vessel measurements and processes are applied, including the second-order local statistical information contained in the Hessian matrix and the matched filter applied in the direction of the largest curvature. Edge suppression reduces sharp brightness changes at finger borders and image contrast is non-linearly increased through fuzzy enhancement. A novel probabilistic fuzzy C-means clustering algorithm is used to derive vessels from the surrounding tissue regions using spatial information in the membership function. Therefore, as shown experimentally, better segmentation and classification rates than the standard C-means algorithm is achieved using the primitive set of features. Moreover, the segmentation results are validated using two cluster-based functions: partition coefficient and partition entropy. Over-segmentation conditions are handled using a two-stage morphological post-processing. Morphological majority filter smoothes the vessel contours and removes small isolated regions which have been misclassified as vessels. Morphological reconstruction is used to obtain outlier-free vessel pattern. The proposed algorithm is evaluated both in real and artificially created images and under different noise types and signal-to-noise ratios, giving excellent segmentation accuracy in the main vessels, even in case where strong artificial noise is used to distort the images. Furthermore, the algorithm can readily be applied in many image enhancement and segmentation applications.

Nearest neighbor (NN) rule is one of the simplest and the most important methods in pattern recognition. In this paper, we propose a kernel difference-weighted k-nearest neighbor (KDF-KNN) method for pattern classification. The proposed method defines the weighted KNN rule as a constrained optimization problem, and we then propose an efficient solution to compute the weights of different nearest neighbors. Unlike traditional distance-weighted KNN which assigns different weights to the nearest neighbors according to the distance to the unclassified sample, difference-weighted KNN weighs the nearest neighbors by using both the correlation of the differences between the unclassified sample and its nearest neighbors. To take into account the effective nonlinear structure information, we further extend difference-weighted KNN to its kernel version KDF-KNN. Our experimental results indicate that KDF-WKNN is much better than the original KNN and the distance-weighted KNN methods, and is comparable to or better than several state-of-the-art methods in terms of classification accuracy.