Shefali Srivastava

BeamAtt: Generating Medical Diagnosis from Chest X-Rays using Sampling Based Intelligence

Aman Sawarn, Shefali Srivastava, Monika Gupta, Smriti Srivastava

Medical Imaging coupled with Image Captioning is fuelling the possibilities of generating accurate medical reports with minimal human intervention. In economically downtrodden nations, this produces opportunity for the poor to acquire world-class treatment from around the globe with an efficient time to market. Chest X-Ray images are integral to the task of diagnosis and treatment of respiratory problems. We propose BeamAtt: an end-to-end deep CNN-RNN based encoder-decoder framework that incorporates spatial visual attention to generate a terse diagnosis from Chest X-ray films. We choose to use a GRU RNN decoder. To boost performance over state-of-the-art methods with complex architectures, we employ Sampling based techniques along with Beam Search Optimisation while generating inferences and argue that a simpler framework with intelligent optimisation is able to successfully achieve higher performance metrics. We show how vivid attention plots can provide deep insight into the region of the image on which the network concentrates to generate a word token. We compare our model with recent prior art using standard evaluation metrics BLEU-1/2/3/4 and ROUGE-L and demonstrate superiority of the proposed method. BeamAtt achieves a BLEU-1 score of 0.56 and CIDEr score of 2.077 which is a significant boost in performance over contemporary solutions.

Neural Network based Routing Protocol for Opportunistic Networks with Intelligent Water Drop Optimization

Shefali Srivastava*, Shubham Kumaram*, Deepak Kumar Sharma

Opportunistic Networks (OppNets) is a system of wirelessly connected nodes in a varying network topology. Routing in OppNets is a challenge. To overcome the problem of routing, an intelligent dynamic strategy to select the next best node for forwarding a message is required. This paper proposes an intelligent routing mechanism based on Intelligent Water Drop (IWD) Algorithm which is used in tandem with Neural Networks (NNs) as an optimization technique to solve the problem of routing in such networks. The nature–inspired IWD algorithm provides robustness, whereas the neural network base of the algorithm helps it to make intelligent routing decisions. The weights in the Neural Network model are calculated by IWD Algorithm using training data consisting of inputs that are characteristic parameters of nodes, such as buffer space, number of successful deliveries and energy levels along with transitive parameters such as delivery probabilities. The proposed protocol Intelligent Water Drop Neural Network (IWDNN) is compared with other protocols that use similar ideologies such as MLProph, K‐nearest neighbour classification based routing protocol (KNNR), Cognitive Routing Protocol for Opportunistic Network (CRPO), and Inheritance Inspired Context Aware Routing Protocol (IICAR), as well as the standard protocol Prophet. IWDNN is shown to outperform all other protocols with an average message delivery ratio of 60%, which is a significant improvement of over 10% in comparison to other similarly conceived algorithms. It has one of the lowest latency among the protocols studied, in a range of 3000 to 4000 s, and incurs comparably low overhead costs in the range of 15 to 30. The drop ratios are one of the lowest, staying near six and approaching zero as buffer size is increased. Average amount of time a message stayed in the buffer was the lowest, with a mean of 1600 s.

Matching Disparate Image Pairs using Shape Aware Conv-Nets

Shefali Srivastava*, Abhimanyu Chopra*, Arun C.S. Kumar, Suchendra M. Bhandarkar, Deepak Sharma

An end-to-end trainable ConvNet architecture, that learns to harness the power of shape representation for matching disparate image pairs, is proposed. Disparate image pairs are deemed those that exhibit strong affine variations in scale, viewpoint and projection parameters accompanied by the presence of partial or complete occlusion of objects and extreme variations in ambient illumination. Under these challenging conditions, neither local nor global feature-based image matching methods, when used in isolation, have been observed to be effective. The proposed correspondence determination scheme for matching disparate images exploits high-level shape cues that are derived from low-level local feature descriptors, thus combining the best of both worlds. A graph-based representation for the disparate image pair is generated by constructing an affinity matrix that embeds the distances between feature points in two images, thus modeling the correspondence determination problem as one of graph matching. The eigen-spectrum of the affinity matrix, i.e., the learned global shape representation, is then used to further regress the transformation or homography that defines the correspondence between the source image and target image. The proposed scheme is shown to yield state-of-the-art results for both, coarse-level shape matching as well as fine point-wise correspondence determination.

Deep Spectral Correspondence for Matching Disparate Image Pairs

Arun C.S. Kumar*, Shefali Srivastava*, Anirban Mukhopadhyay, Suchendra M. Bhandarkar

A novel, non-learning-based, saliency-aware, shape-cognizant correspondence determination technique is proposed for matching image pairs that are significantly disparate in nature. Images in the real world often exhibit high degrees of variation in scale, orientation, viewpoint, illumination and affine projection parameters, and are often accompanied by the presence of textureless regions and complete or partial occlusion of scene objects. The above conditions confound most correspondence determination techniques by rendering impractical the use of global contour-based descriptors or local pixel-level features for establishing correspondence. The proposed deep spectral correspondence (DSC) determination scheme harnesses the representational power of local feature descriptors to derive a complex high-level global shape representation for matching disparate images. The proposed scheme reasons about correspondence between disparate images using high-level global shape cues derived from low-level local feature descriptors. Consequently, the proposed scheme enjoys the best of both worlds, i.e., a high degree of invariance to affine parameters such as scale, orientation, viewpoint, illumination afforded by the global shape cues and robustness to occlusion provided by the low-level feature descriptors. While the shape-based component within the proposed scheme infers what to look for, an additional saliency-based component dictates where to look at thereby tackling the noisy correspondences arising from the presence of textureless regions and complex backgrounds. In the proposed scheme, a joint image graph is constructed using distances computed between interest points in the appearance (i.e., image) space. Eigen Spectral decomposition of the joint image graph allows for reasoning about shape similarity to be performed jointly, in the appearance space and eigenspace. Furthermore, a new benchmark dataset consisting of disparate image pairs with extremely challenging variations in scale, orientation, viewpoint, illumination and affine projection parameters and characterized by the presence of complete or partial occlusion of objects, is introduced. The proposed dataset is supplemented with ground truth interest point annotations and is the largest and most comprehensive amongst publicly available image datasets pertaining to the problem of disparate image matching. The proposed scheme yields state-of-the-art performance in the case of both, coarse-grained shape-based correspondence determination as well as fine-grained point-wise correspondence determination on two existing challenging datasets as well as the newly introduced dataset.

Biometric authentication using local sub- space adaptive histogram equalization

Gopal Chaudhary, Shefali Srivastava, Smriti Srivastava

In biometrics authentication systems, such as palmprint recognition, fingerprint recognition, dorsal hand vein recognition and palm vein recognition etc., image enhancement plays a crucial role for most of the low resolution image samples. In this work, a novel adaptive histogram equalization (AHE) variant is proposed as effective area-AHE (EA-AHE) with weights. Here, global adaptive histogram equalization is improved using a local AHE technique by varying the effective area with different effective weights. The method is found to improve the biometric authentication identification rate as compared to the typical AHE. To validate the proposed algorithm, IITD palmprint databases of left and right hand are used in the simulations. Finally, it is validated through results that proposed technique is superior to the existing ones.

Information Fusion in Animal Biometric Identification

Gopal Chaudhary, Smriti Srivastava, Saurabh Bhardwaj, Shefali Srivastava

This work presents the application of biometrics in animal identification, which is a highly researched topic in human recognition. Here, our analysis presents the identification of zebra in their natural habitat. All the techniques are tested on 824 Plains zebra images captured at Ol’Pejeta conservancy in Laikipia, Kenya. We have used coat strips as a biometric identifier which is unique in nature. To improve the performance of identification, information fusion of coat strips can be taken place from many points in zebra skin such as near legs, stomach and neck. Here two region near stomach (flake) and first limb (leg) is cropped from the textural pattern of strips of zebra is used in feature extraction. GMF, AAD, mean, and eigenface feature extraction methods are applied on flake and limb ROI of zebra. Then a novel image enhancement method: difference subplane adaptive histogram equalization is applied to improve the identification rate. Our technique is based on information fusion in fusing the score from stomach (flake) and first limb (leg) region. For this, sum, product, frank T-norm, and Hamacher T-norm rules are applied to validate the identification results. Information fusion improves the identification results from the previous reported results from eigenface, CO-1 algorithm, and stripecodes. The improvement in results verifies the success of our approach of information fusion using score level fusion.