Pulmonary fissure detection is an important step for lung lobe segmentation which is necessary for accurate diagnostics and surgical planning. Automatic detection of fissures in CT images is a challenging task due to varying intensity, pathological deformation and noisy acquisitions. In this paper, we propose a novel fissure line detection technique using eigen analysis of the hessian matrix and an exhaustive sample consensus (ESAC) based line fitting in small overlapping windows. The idea behind using the line fitting technique is that the fissure line appears as piece-wise linear segment in a small window. As opposed to RANSAC, the point selection mechanism in the proposed method chooses all combination of data points exhaustively. This approach reduces the possibility of missing the possible candidate points for a fissure line. Our main contribution lies in detection of the fissure line without using any training data as well as any template matching model. The performance of our method is validated on the publicly available LOLA11 database. Comparisons with some existing approaches on this database indicate the advantage of the proposed solution.
In this paper, we propose a novel framework for anomaly detection and categorization in crowded scenes via location aware hierarchical cell-based (LAHCB) approach. Input video frames are split into spatio-temporal volumes (STVs). Low-level histogram of optical flow orientation and motion magnitude features are extracted from selective STVs (SSTVs) at three levels of hierarchy achieving a coarse-to-the-fine localization. We use one-class SVM (OCSVM) classifier to model normal events and detect anomaly. The proposed method consists of two parts: (1) Global Analysis for frame-level detection; (2) Local Analysis for pixel-level localization. For global analysis, we adopt computationally less expensive model that uses only coarser level. Further, we use this information to achieve localization. The computational efficiency lies in faster online testing since storage and time overhead is due to offline feature extraction and classification at different levels in the hierarchical structure during training phase. Unlike existing methods, the proposed approach omits pixel level feature computation and background modeling. The addition of location aware concept detects abnormal behaviour in an unexpected region. We demonstrate the performance of the proposed method on the UCSD and UMN datasets. We achieve AUC of 86.16% and 86.3% on Ped1 and Ped2, comparable with state-of-the-art methods.
This paper attempts to solve an major problem of rigorous subject training in BCI based robotics. A hybrid Brain computer interface has been established here to mentally guide a robot arm without any motor commands generated in mind. Spontaneous N200 response of human brain as a part of motion onset visual evoked potential (m VEP) is used here to detect the desired object in the environment and use of Steady state visual evoked potential (SSVEP) provides the guidance to a bedside robotic arm to reach that target object. Electroencephalographic response of ten such subjects has been used here to evaluate the efficacy of the proposed system. Recorded EEG response goes through a sequential operation of preprocessing, feature extraction and classification. For detecting N200 response of the brain, a novel EKF-Particle filter induced Neural Network classifier is also proposed which essentially outperforms the other existing classifier for N200 detection. Performance analysis of other classifiers has been given here for comparison.
Due to high variability in writing style of different individuals, non-centered and non-uniformly scaled optical characters are very difficult to recognize. Several techniques are proposed in-order to solve the recognition problem. In this work, we highlight that the performance of optical character classifiers which are based on the deep learning framework can be improved through feature-map alignment. Here, we have used spatial transformer network to align the feature maps of a convolutional neural network model which is proposed for the classification problem. We demonstrate that with the proposed framework not only the slight transformed versions which are usually considered in the conventional datasets can be classified with high accuracy, but also highly non-uniform in scale characters can also be fairly recognized with quite higher accuracy. We evaluate our proposed model on CMATERdb 3.1.3 database which consists of isolated Bangla handwritten compound characters and our model obtained 97.86 % recognition accuracy in the original database and 96.34 % on various rotated data in training and testing.
Handwritten pen or finger based inputs are very common, especially with hand-held devices like mobile phones or PDAs. Apart from the content, there is the writing style information embedded in to the handwritten input. They are useful as personalized command gestures, passwords or as authentication supplements. Since the input is assumed to be written by a finger on the touch screen of the hand-held device, the input considered is a trajectory over the (x; y; t) space where x and y are the spacial and t is the time co-ordinate. The hand held device can be held in a variety of ways while giving the input, and size of the input may also vary. Normalizing with respect to rotation is an error prone process. Also, the other important limitation is that the number of training examples could be very few. The paper presents a novel style preserving TRS (translation, rotation and scale) invariant representation scheme for the 3D trajectory data that use 2D Zernike moments only and is called the time varying Zernike moments (TVZMs). For this, we present a 2D representation of the 3D function and show that it is lossless. The proposed representation scheme is sensitive to both content and to the writing style. Experimentally it is shown that 2D TVZMs are superior than 3D Zernike moments which possess same invariance properties. Also a comparison is drawn with a recent online signature representation scheme. Hence, it may be a suitable one for personalized command gesture recognition.
Occupants and their actions play major roles in building energy management as reported by previous studies, which involves finding the optimal schedule of user actions, under a given physical context, in order to minimize their dissatisfaction. However, comparison and performance analysis of various optimizers, for the concerned problem, have not been studied previously, which is essential to gain insight into the underlying characteristics of the problem. In this work, the performance of four popular and contemporary multi-objective optimization algorithms viz. DEMO, NSGA-II, NSGA-III, and Θ-DEA, for estimating the optimal schedule has been analyzed in terms of their abilities to find minimal average indoor conditions' to discover more number of alternative trade-off solutions (flexibility) and to promptly converge to a smaller minimal net dissatisfaction value (speed of convergence). Results show that NSGA-II has slightly better capabilities than NSGA-III and Θ-DEA, but it clearly outperforms DEMO. The recently developed population dynamics indicators are also applied to support the observed features of the optimizers. The proposed analyzing paradigm can also be used when the optimization problem is extended to include several other objectives.
The process of determining which disease or condition explains a persons symptoms and signs can be very complicated and may be inaccurate in some cases. The general belief is that diagnosing diseases relies on doctors keen intuition, rich experience and professional equipment. In this work, we employ ideas from recent advances in plantar pressure research and from the powerful capacity of the convolutional neural network for learning representations. Here, we propose a model using convolutional neural network based on plantar pressure for medical diagnosis. Our model learns a network that maps plantar pressure data to its corresponding medical diagnostic label. We then apply our model to make the medical diagnosis on datasets we collected from cooperative hospital and achieve an accuracy of 98.36%. We demonstrate that the model base on the convolutional neural network is competitive in medical diagnosis.
Automatic facial expression recognition is one of the most recently topic in aspect of behaviour analysis and human computer interaction (HCI). Difficulty with facial expression recognition system is to implement generic model. Same facial expression may vary across humans, even this is true for the same person when the expression is displayed in different situations. This paper proposed the local image descriptor that extracts the histogram of second order gradients (HSOG), which capture the local curvatures of differential geometry. The shape index is computed from the curvatures and its different values correspond to different shapes. In case of facial expression recognition using full face images, if any portion of the face image is distorted, it may reflect on the recognition performance. Humans have the capability to recognize faces even by looking at some parts of the face. An attempt has been made to replicate the same on machines by only considering some of the informative regions of the face like eyes, nose, lip and forehead. Facial expression recognition experiments have been performed on some benchmark databases, Better recognition rates were achieved compared to other existing approaches.
This paper present a new handwritten English word database namely, ISIHWD for off-line handwritten word recognition and writer identification. It is a collection of 62 English legal amount words used in Indian bank cheque written by 105 different writers. The database consist of 31124 handwritten words split into a training set of 24924 words and a test set of 6200 words to a provide common platform in writer independent word recognition. The database is evaluated by deep convolution neural network based feature with SVM classifier for writer independent handwritten word recognition and text dependent writer identification based on single word. Average word recognition and writer identification accuracy are achieved as 95.35% and 85.70% respectively.
This paper presents a novel image inpainting method for the completion of image structures in digital circuit images. Here we have proposed a set of geometric patch transformations in order to facilitate searching good candidate patches. Furthermore, we incorporate these transformations in an objective function that comprises both color-based approach and gradient domain method in a single framework to expedite global optimization. The motivation of this approach is to solve the problem of propagating geometric structures smoothly inward the target region. Our image inpainting process consists of two core steps: search and voting. We alternate these two steps until a suitable convergence criterion is satisfied. We repeat the process in a multiscale approach, starting from coarsest scale and ending at finest scale. The proposed method is tested on some circuit images, and the results are compared with some of the existing methods to demonstrate the efficacy and superiority of the proposed method.
This paper presents PolicyGAN, a policy gradient paradigm for training Generative Adversarial Networks that views the generator as an image generation neural agent which is rewarded by another neural agent, termed as the discriminator. Rewards are higher for samples near the original data manifold. In PolicyGAN, only reward signal from the output of the discriminator is used for updating the generator network using policy gradient. This obviates the need for gradient signal to flow through the discriminator for training the generator; an intrinsic property of original GAN formulation. Given the inherent difficulty of training adversarial models, and low convergence speed of policy gradient, training GANs using policy gradient is a non-trivial problem and requires deep study. Currently GANs have used only differentiable discriminators for training. Policy-GAN germinates the possibility of using a wide variety of non-differentiable discriminator networks for training GANs, something which was not possible with the original GAN framework. Another advantage of using policy gradient is that now the generator need not produce deterministic samples, but can generate a probability distribution from which samples can be taken. PolicyGAN thus paves the path to use a variety of probabilistic models.
Segmenting a natural image is a complex task. Different semantic units may share similar visual features. On the other hand, such features can have variations even within a single unit. Proposed methodology relies on colour cohesion and spatial relationship between the components with cohesive colour. At first image colour space is clustered to map the original colour to a reduced set. Number of cluster is automatically detected by analyzing the intensity histograms of the colour channels. Based on the similarity in terms of mapped colours, pixels are grouped. Subsequently, the spatial inclusiveness criteria is considered to merge the pixels groups where one group is contained within another. Finally, an attempt is made to merge the adjacent regions based on colour gradient. Colour cohesion is conceptualized by the process of colour space clustering, grouping of pixels in terms of colour similarity and region merging based on colour gradient. The spatial criteria is taken into account in terms of spatial inclusiveness at intermediate level and adjacency at final stage. Proposed methodology is tested on Berkley segmentation dataset. Performance comparison with few other methodologies indicates the effectiveness of proposed methodology.
India being an agricultural land, the monsoon governs the pulse of life for its mankind along with the flora-fauna existing in the subcontinent. Active and break spells are two important phases of monsoon which determine the overall quality and quantity of monsoon for a year. Prediction of active and break spells of monsoon is of high significance owing to planning of proper strategy for best utilization of monsoon phases, which assists in economic development of the country. In this article we propose a classification based scheme for the prediction of spells of the Indian summer monsoon. We have used linear discriminant analysis based classifier to classify the monsoon days into into dry, normal, or wet days. Classifying monsoon at daily scale makes it more challenging as uncertainty and variability are more higher as compared to monthly or yearly monsoon prediction. Climatic variables are selected of meteorology importance as input features to the classifier for obtaining higher accuracy of classification. Proposed method initiates with greedy feature selection, followed by linear discriminant analysis based classification of daily monsoon. Assembling the classified monsoon days lead to the determination of active or break spells during the period. Classification results are observed to be close to an acceptable performance, owing to the complexity of the phenomenon. The spells are predicted with comparable accuracy. The accuracy of prediction of the break spells are observed to be superior to that of the active spells.
Protein-protein interaction network(PPI) is a biological network which represents interaction between proteins and these interactions change over time. Our aim is to study how these networks change from normal state to disease state. The normal network transforms into a disease network by new proteins joining the network or existing proteins leaving the network. To model this process, Graph kernel measures are used. By comparing pairs of graphs with different arrival sequences of nodes, we try to discover the arrival sequence that produces a smooth transition during transformation of a normal network into a disease network. Also the nodes that produce significant changes in the network are studied from the perspective of influential nodes theory in social networks. The network evolution is modelled using Duplication-mutation with complementarity (DMC) model. Results at a global level matching of network properties are satisfactory. But the actual matching of links between the simulated network and the original is poor and this issue needs to be investigated further.
The ability to spot a few known characters or symbols allows the linguists and historians to guess the era in which an inscription was made. Manual spotting of these characters proves to be very laborious and error-prone. Hence, automation in character spotting has evolved in recent time, which has its own challenges due to natural wear and tear of inscriptions through aging. In this paper, we present a computationally efficient technique for character spotting using certain concepts from rough set theory. After image binarization, we compute various attributes for the isolated symbols within the ambit of rough set. In order to spot a symbol in the inscription, the corresponding attribute set for the query symbol is matched with that of the inscribed symbols. We provide the details of the method in this paper and show that the symbols or characters can be quite accurately spotted in the inscription.
In the context of video surveillance, detecting violence is an important task. This work presents a deep neural network based novel video classification to label violence/non-violence classes. First of all, frame-level descriptors are constructed based on optical flow and variants of Weber Local Descriptor. A novel scheme is presented to summarize the frame-level descriptors in to the video-level descriptor. An architecture for deep neural network with four hidden layers is proposed to classify a video as violent or non-violent. Proposed method is tested on two benchmark datasets. Comparison of performance with state-of-the-art systems establishes the superiority of the proposed method.
Hemodynamic analyses of cerebrovasculature help in clinical diagnosis of various cerebrovascular diseases. Individual vascular geometry influence various hemodynamic factors as well as one's cerebrovascular health. In this work we present a new 3-D digital flow algorithm for quick simulation of hypothetical fluid flow in the mathematically in generated approximate cerebrovascular phantoms. The algorithm starts from a user initiated flow start point and iteratively wets the object region until convergence or a pre-defined end point is reached. We have also developed a customized 2-D/3-D user interface for initiation of fluid flow. After the termination of the algorithm, it delineates flow direction, flow velocity at each of the point of the object. The experimentation is done on several mathematically generated phantoms.
The generalized pseudo-Jacobi-Fourier moments (G-PJFMs) which are orthogonal and rotation invariant are characterized by a free parameter and the classical pseudo-Jacobi-Fourier moments (PJFMs) are the special case of G-PJFMs. In this paper, we have proposed the multi-channel G-PJFMs (MG-PJFMs) using the multi-channel framework and analyzed the effect of the free parameter on color image reconstruction and color object recognition under various geometric transformations and noisy conditions for G-PJFMs. The detailed experimental analysis provides the optimum values of the free parameter for these image analysis tasks.
In this article parametric curve based a new image encryption method is proposed. Here, to obtain the encrypted image, sine curves (siτι(a, w, Φ, t) = a * siτι(Φ + 2ιτwt)) have used in both confusion and diffusion phases. In different steps, different sine curves have used, with varying the parameters and this provide more security to the proposed method. The result of the proposed method, performance under different statistical analysis and sensitivity analysis established that proposed method provides a secure and efficient way for image encryption.
Speech Synthesis (SS) and Voice Conversion (VC) methods provides a great risk for Automatic Speaker Verification (ASV) system. In this paper, we tried to find the difference between natural and spoofed speech signals using Teager Energy Operator-based Energy Separation Algorithm (TEO-ESA). Here, we exploit the contribution of Amplitude Envelope (AE) and Instantaneous Frequency (IF) in each narrowband filtered signals energy via ESA to capture possible changes in a temporal and spectral envelope of the synthetic speech signal generated by the machines as opposed to natural signals. Furthermore, IF was used for classification of natural vs. spoof speech with Gaussian Mixture Model (GMM) as a classifier. These findings may assist to distinguish these two speeches and provide an aid to alleviate possible impostor attacks in voice biometrics. The experiments are done on ASV Spoof 2015 Challenge database. We have compared proposed Energy Separation Algorithm-Instantaneous Frequency Cosine Coefficients (ESA-IFCC) with Mel Frequency Cepstral Coefficients (MFCC) features. On the development set, MFCC alone gave an Equal Error Rate (EER) of (6.98 %) and ESA-IFCC gave (5.43 %) with 13-D static features. With score-level fusion of MFCC and ESA-IFCC EER reduced to 3.45 % on static feature vector. The EER decreases further to 2.01 % and 1.89 % for Δ and ΔΔ features. On evaluation set, the overall average error rate for known and unknown attacks was 6.79 % for ESA-IFCC and was significantly better than the MFCC (9.15 %) and their score-level fused EER (7.16 %).
