Search

This paper proposes generating Indian Sign Language from sentences using sentence preprocessing and Generative Adversarial Networks (GANs). The proposed method focuses on converting sentences to brief notations in the form of glosses and generating synthetic video frames for each gloss using a Generative Adversarial Network (GAN) architecture, thus resulting in a video representation for the input sentence. The glosses are mapped to a corresponding minimal skeletal representation, which is used as an input and a base image for the GAN model. The GAN outputs the corresponding video frames, which are further processed to produce a video. The dataset consists of video representations of words present in the Indian Sign Language. The proposed paper measures the performance of the method by using the SSIM index and a side-by-side comparison of real and synthetically generated video frames.

IndoorNet is an approach to generate 3D layouts for indoor scenes. By employing a fully convolutional network (FCN), it predicts the layout of the room using only a single 360 degree panorama image. The FCN uses geometrical features, i.e., vanishing points along with the panoramic image for better contextual understanding. Further, we optimize the boundary map and the corner map to improve their accuracy. We then use these improved maps to generate the 3D layout of the room. As compared to the state of the art, our network architecture is faster and lighter, with improved results.

Language is considered an integral part of human communication. Sometimes, a non-native speaker who is as fluent as the native speaker may not be comprehended suitably because of dissimilitude in accents. Also, detecting a person’s accent itself poses an arduous task due to subtle heterogeneity in pronunciation. In order to curtail these communication barriers arising because of accents distinctness, we have proposed a system to detect and convert speech which can conveniently differentiate one accent from another which has achieved an accuracy of 68.67%. The main motivation was solving the difficulty of Indians to understand the foreign accents and of foreigners to understand the Indian accent. The scope of this project would be limited to the English language. Detection of accent would involve identifying the speaker’s native language. We introduced a novel method of augmenting the traditionally exercised techniques by using Convolutional Neural Network (CNN) fused with Deep Neural Network (DNN) along with Recurrent Neural Networks (RNN) to subsequently enhance the accuracy. For accent conversion, audio features mainly consist of Mel-Frequency Cepstral Coefficients (MFCCs), Aperiodicity (AP) and Fundamental Frequency (F0). These features once extracted are processed by a Generative Adversarial Network (GAN) to convert source accent to target accent.

The objective of this project is to automate the detection of Diabetic Retinopathy and evaluate the severity with high efficiency, through an overall feasible approach. This project explores the use of various Convolutional Neural Network Architectures on images from the dataset after being subjected to appropriate image processing techniques like local average color subtraction to aid in highlighting the germane features from a fundoscopy, thereby augmenting the detection and evaluation process of Diabetic Retinopathy as well as serve as an expert guidance system for practitioners around the world.

With the increasing demands on 3D applications and the easy capturing of 2D images nowadays, building 3D models from 2D images receives much attention in the past few years. 3D modeling is widely used in several fields3D graphics in computer games, software architecture models and 3D printing. 3D models represent a physical body using a collection of points in 3D space, connected by various geometric entities such as triangles, lines, curved surfaces, etc. 3D modeling is the process of developing a mathematical representation of any three-dimensional surface of an object. Today, 3D models are used in a wide variety of fields. The engineering community uses them as designs of new devices, vehicles and structures as well as a host of other uses. A variety of machine learning algorithms are being studied and implemented to find or estimate the depth information which is unavailable in conventional 2D image. We apply Computer Vision algorithm considering aspect of binocular disparity where we use 2 images of same scene captured from different viewpoints. Then we obtain depth of the object and further construct depth map. After mapping the points from depth maps of various images captured we apply correct texturing to obtain full 3D model of the object.capturing one eye‟s view, and depth information is computed using binocular disparity. Here we focus only on binocular and multi-ocular images as input. The two or more input images could be taken either by multiple fixed cameras located at different viewing angles or by a single camera with moving objects in the scenes. A three-dimensional (3D) visualization enables consumers to interact with products and creates a sense of being in a simulated real world. As the consumer gets a real view of the products they tend to get attracted towards the product thus increasing the sale. It gives us an edge over other competitors as 3D visualization is different and it stands out from others. It also makes shopping more convenient and easy for the customers. In this paper we focus only on binocular and multiocular images as input. We study computer vision algorithm, binocular disparity, silhouette and visual hull. In computer vision algorithm, SURF and ORB features descriptors are used to extract information from images. Binocular disparity uses 2 images of the same scene from different viewpoints. In silhouette the object is separated from the background and silhouette cones are formed. Intersection of silhouette cones is called visual hulls. General Terms 2D images, Computer Vision, 3D reconstruction

aging is one of the most challenging task in image processing and is commonly used in many areas. This paper consists of compositional method which will represent faces of different age groups. The representation of faces in different age group is done hierarchically i.e. using And-Or graph, in which the And nodes will decompose the face into different components (e.g. wrinkle, hair) for crucial age perception and the Or node will represent the diversities of faces. The graph is then represented using Markov chain. The diversities or the uncertainties of the faces are learnt from the large database which consists of large number of images. There are two criteria for evaluating the result the aging simulation and one of them is the accuracy of the simulation i.e. whether the perceived image belongs to a particular age group, and second is the preservation of the identity i.e. whether the face, that is retrieved after simulation process is preserving the identity of the person or not. The statistical analysis of these two above mentioned criteria will decide the performance of the aging simulation. Keywords-Or Graph, Aging modeling, ANOVA.

Object tracking is the key feature of video analysis. In this paper, we present a method to track object in a multi-sport field. Mean shift algorithm is a kernel tracking approach which tracks on the basis of color and shape of the target object. Due to constant motion in a playground, it is difficult to track the same body because of changing color and direction. Hence we introduce a novel method, Incremental Discriminative Color Object Tracking to track entities on a sports ground. We have proposed to modify certain elements of the original Mean Shift algorithm so as to track entities in video streams with changing color, shape and direction. The trajectory formed from this method enables us to calculate the distance covered by an entity in the field.

Baby Mentor is a system developed to learn and teach babies through images. Through intuitive observations and studies it is shown that using a visual approach i.e. images, videos is an effective method for teaching babies and young children. So for this purpose we present a system that automatically recognizes and generates automatic real time for objects such as toys, picture cards or flashcards that are commonly used by children. The method begins with detecting and extracting features from image for which Scale Invariant Feature Transform (SIFT) algorithm is used. SIFT is scale, rotation and transformation invariant thereby making it easier to identify objects even if it is rotated or placed in dim light. The bag-of-visual-words model (BOVW) is then used to accumulate, cluster and segregate the features of the image generated by the SIFT algorithm, into their respective classes. This helps in generating vocabulary needed to generate the natural language description. Finally the text generated is spoken out through a speaker or microphone. This will help babies especially those suffering from autism in their learning and communication process by helping them identify and understand common world objects with the help of the text and speech generated. Besides this the entire process would require minimal human assistance thus making it a convenient and easy-to-use application.

With the advent of new and improved technology in the market, it has become extremely important to gather information and create user friendly applications. Be it for security purposes or tracking purposes, monitoring human activities to create a human gesture-friendly interface is now a major branch of research. We hope to present one such method to observe and interpret human activities. This paper describes methods to obtain human activity records (in the form of images or videos) and perform suitable actions on these records to extract useful information out of them. Background subtraction is the initial one of the most important procedures performed on the image or video. Next comes preprocessing the record to improve quality. Then finally comes Skeletonisation. We focus on Background Subtraction method using the image pixels and Skeletonisation in this paper and provide the basic steps and strategies regarding it.

Unmanned aerial vehicles (UAV), also referred to as drones, are a growing field in computer science with applications in military systems, delivery services, emergency relief and evacuation. One of the primary obstructions to the allowance of UAV journeys over populated areas is the lack of sophisticated automated systems that detect drone landing sites. In this paper, we propose a landing area detection system using machine learning and image processing. This system compares the suitability of various features (RGB Color Model, HSV Color Model, LBP, Edge Density) in determining a suitable drop-off point. Classification on these features has been carried out using Support Vector Machines (Linear, Polynomial and RBF Kernel).

In this paper, an unusual form of image registration is proposed which focuses upon using various transforms for fast and accurate image registration. Image registration is the process of transforming different sets of data into one coordinate system. The data set can be a set of photographs, data from various sensors, from different times, or from different viewpoints. The applications of image registration are in the field of computer vision, medical imaging, military automatic target recognition, and in analyzing images and data from satellites. The proposed technique works upon satellite images. It tries to find out area of interest by comparing the unregistered image with source image and finding the part that has highest similarity matching. The paper mainly works on the concept of seeking water or land in a stored image. The proposed technique uses different transforms like Discrete Cosine Transform, Discrete Wavelet Transform, HAAR Transform and Walsh transform to achieve accurate image registration. The paper also focuses upon using normalized cross correlation as an area based technique of image registration for the purpose of comparison. The proposed algorithm is worked over various sizes of satellite images such as 256×256, 1024×1024 etc. The root mean square error is used as similarity measure. The experiment results show that the proposed algorithm can successfully register the template and can also process local distortion in high-resolution satellite images.

Image registration is the process of geometrically aligning one image to another image of the same scene taken from different viewpoints or by different sensors. It is a fundamental image processing technique and is very useful in integrating information from different sensors, finding changes in images taken at different times and inferring three-dimensional information from stereo images. Image registration can be done by using two matching method: transform based methods and correlation based methods. When image registration is done using correlation based methods like normalized cross correlation, the results are slow. They are also computationally complex and sensitive to the image intensity changes which are caused by noise and varying illumination. In this paper, an unusual form of image registration is proposed which focuses upon using various transforms for fast and accurate image registration. The data set can be a set of photographs, data from various sensors, from different times, or from different viewpoints. The applications of image registration are in the field of computer vision, medical imaging, military automatic target recognition, and in analyzing images and data from satellites. The proposed technique works on satellite images. It tries to find out area of interest by comparing the unregistered image with source image and finding the part that has highest similarity matching. The paper mainly works on the concept of seeking water or land in the stored image. The proposed technique uses different transforms like Discrete Cosine Transform, Discrete Wavelet Transform, HAAR Transform and Walsh transform to achieve accurate image registration. The paper also focuses upon using normalized cross correlation as an area based technique of image registration for the purpose of comparison. The root mean square error is used as similarity measure. Experimental results show that the proposed algorithm can successfully register the template and can also process local distortion in high-resolution satellite images.