Balraj Ashwath

Balraj Ashwath

Senior ML Engineer | LLMs, Agentic AI, Computer Vision

Publications

Towards Practical and Efficient High-Resolution HDR Deghosting with CNN

Published in European Conference on Computer Vision (ECCV), 2020

Generating High Dynamic Range (HDR) image in the presence of camera and object motion is a tedious task. If uncorrected, these motions will manifest as ghosting artifacts in the fused HDR image. On one end of the spectrum, there exist methods that generate highquality results that are computationally demanding and too slow. On the other end, there are few faster methods that produce unsatisfactory results. With ever increasing sensor/display resolution, currently we are very much in need of faster methods that produce high-quality images. In this paper, we present a deep neural network based approach to generate high-quality ghost-free HDR for high-resolution images. Our proposed method is fast and fuses a sequence of three high-resolution images (16-megapixel resolution) in about 10 seconds. Through experiments and ablations, on different publicly available datasets, we show that the proposed method achieves state-of-the-art performance in terms of accuracy and speed.

Download here

Dense Refinement Residual Network for Road Extraction From Aerial Imagery Data

Published in IEEE Access, 2019

Extraction of roads from high-resolution aerial images with a high degree of accuracy is a prerequisite in various applications. In aerial images, road pixels and background pixels are generally in the ratio of ones-to-tens, which implies a class imbalance problem. Existing semantic segmentation architectures generally do well in road dominated cases but fail in background-dominated scenarios. This paper proposes a dense refinement residual network (DRR Net) for semantic segmentation of aerial imagery data. The proposed semantic segmentation architecture is composed of multiple DRR modules for the extraction of diversified roads alleviating the class imbalance problem. Each module of the proposed architecture utilizes dense convolutions at various scales only in the encoder for feature learning. Residual connections in each module of the proposed architecture provide the guided learning path by propagating the combined features to subsequent DRR modules. Segmentation maps undergo various levels of refinement based on the number of DRR modules utilized in the architecture. To emphasize more on small object instances, the proposed architecture has been trained with a composite loss function. The qualitative and quantitative results are reported by utilizing the Massachusetts roads dataset. The experimental results report that the proposed architecture provides better results as compared to other recent architectures

Download here