arXiv subjects

Kaushik Das

Publications and source records attributed to Kaushik Das.

23 records · Page 2Linked to original sources

arXiv2020

Convolutional Recurrent Residual U-Net Embedded with Attention Mechanism and Focal Tversky Loss Function for Cancerous Nuclei Detection

Since the beginning of this decade, CNN has been a very successful tool in the field of Computer Vision tasks.The invention of CNN was inspired from neuroscience and it shares a lot of anatomical similarities with our visual system.Inspired by the anatomyof humanvisual system, wearguethat the existing U-Net architecture can be improvedin many ways. As human visual system uses attention mechanism, we have used attention concatenation in place of normalconcatenation.Although, CNN is purely feed-forward in nature but anatomical evidences show that our brain contains recurrent synapses and they often outnumber feed-forward and top-down connections. Thisfact inspiresus to userecurrent convolution connectionsin place of normalconvolution blocksin U-Net.Thispaper also addressesthe class imbalance issuein the field of medical image analysis. The paperresolvestheproblem of class imbalanceswith the help of state-of-the-art loss functions.Weargue thatourproposed architecturecan be trained end to end with a few training data and it outperforms the other variantsof U-Net.

eess.IV↗

arXiv2020

Implementation of Survivor Detection Strategies Using Drones

Survivors stranded during floods tend to seek refuge on dry land. It is important to search for these survivors and help them reach safety as quickly as possible. The terrain in such situations however, is heavily damaged and restricts the movement of emergency personnel towards these survivors. Therefore, it is advantageous to utilize Unmanned Aerial Vehicles (UAVs) in cooperation with on-ground first responders to aid search and rescue efforts. In this article we demonstrate an implementation and improvement of the weight-based path planning algorithm using an off-the-shelf UAV. The coordinates of the survivor and their heading is reported by an on-ground observer to the UAV to generate a weighted map of the surroundings for exploration. Each coordinate in the map is assigned a weight which dictates the priority of exploration. These waypoints are then sorted on the basis of their weights to arrive at an ordered list for exploration by the UAV. We developed the model in MATLAB, followed by prototyping on Robot Operating System (ROS) using a 3DR Iris quadcopter. We tested the model on an off-the-shelf UAV by utilizing the MAVROS and MAVLINK capabilities of ROS. During the implementation of the algorithm on the UAV, several additional factors such as unreliable GPS signals and limited field of view which could effect the performance of the model were in effect, despite which the algorithm performed fairly well. We compared our model with conventional algorithms described in the literature, and showed that our implementation outperforms them.

cs.RO↗

arXiv2019

Global Optimization of Offshore Wind Farm Collection Systems

A mathematical program for global optimization of the cable layout of Offshore Wind Farms (OWFs) is presented. The model consists on a Mixed Integer Linear Program (MILP). Modern branch-and-cut solvers are able to solve large-scale instances, defined by more than hundred Wind Turbines (WTs), and a reasonable number of Offshore Substations (OSSs). In addition to the MILP model to optimize total cable length or initial investment, a pre-processing strategy is proposed in order to incorporate total electrical power losses into the objective function. High fidelity models are adapted to calculate cables current capacities, spatial currents. The MILP model is embedded in an iterative algorithmic framework, consisting in solving a sequence of problems with increasing size of the search space. The search space is defined as a set of underlying candidate arcs. The applicability of the method is illustrated through 10 case studies of real-world large-scale wind farms. Results show that: (i) feasible points can quickly be obtained in seconds, (ii) points near the global optimum with an imposed maximum tolerance, are calculable in reasonable computational time in the order of hours, and (iii) the proposed method compares favorably against a state-of-the art method available in literature.

math.OC↗

arXiv2018

A Deeper Insight into the UnDEMoN: Unsupervised Deep Network for Depth and Ego-Motion Estimation

This paper presents an unsupervised deep learning framework called UnDEMoN for estimating dense depth map and 6-DoF camera pose information directly from monocular images. The proposed network is trained using unlabeled monocular stereo image pairs and is shown to provide superior performance in depth and ego-motion estimation compared to the existing state-of-the-art. These improvements are achieved by introducing a new objective function that aims to minimize spatial as well as temporal reconstruction losses simultaneously. These losses are defined using bi-linear sampling kernel and penalized using the Charbonnier penalty function. The objective function, thus created, provides robustness to image gradient noises thereby improving the overall estimation accuracy without resorting to any coarse to fine strategies which are currently prevalent in the literature. Another novelty lies in the fact that we combine a disparity-based depth estimation network with a pose estimation network to obtain absolute scale-aware 6 DOF Camera pose and superior depth map. The effectiveness of the proposed approach is demonstrated through performance comparison with the existing supervised and unsupervised methods on the KITTI driving dataset.

cs.CV↗

arXiv2018

UnDEMoN 2.0: Improved Depth and Ego Motion Estimation through Deep Image Sampling

In this paper, we provide an improved version of UnDEMoN model for depth and ego motion estimation from monocular images. The improvement is achieved by combining the standard bi-linear sampler with a deep network based image sampling model (DIS-NET) to provide better image reconstruction capabilities on which the depth estimation accuracy depends in un-supervised learning models. While DIS-NET provides higher order regression and larger input search space, the bi-linear sampler provides geometric constraints necessary for reducing the size of the solution space for an ill-posed problem of this kind. This combination is shown to provide significant improvement in depth and pose estimation accuracy outperforming all existing state-of-the-art methods in this category. In addition, the modified network uses far less number of tunable parameters making it one of the lightest deep network model for depth estimation. The proposed model is labeled as "UnDEMoN 2.0" indicating an improvement over the existing UnDEMoN model. The efficacy of the proposed model is demonstrated through rigorous experimental analysis on the standard KITTI dataset.

cs.CV↗