Avigyan Bhattacharya

Avigyan Bhattacharya

Auto-generated profile, based on publicly available information

Avigyan Bhattacharya

Carnegie Mellon University, Pittsburgh, USA

Research Focus & Metrics

Academic Career Length

4Years (2023-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
15 December 2023
Verified
Conference Paper
Open access
ABAvigyan Bhattacharya
Avigyan Bhattacharya
MSMainak Singha
Mainak Singha
AJAnkit Jha
Ankit Jha
Biplab BanerjeeBiplab Banerjee
Biplab Banerjee
We focus on domain and class generalization problems in analyzing optical remote sensing images, using the large-scale pre-trained vision-language model (VLM), CLIP. While contrastively trained VLMs show impressive zero-shot generalization performance, their effectiveness is limited when dealing with diverse domains during training and testing. Existing prompt learning techniques overlook the importance of incorporating domain and content information into the prompts, which results in a drop in performance while dealing with such multi-domain data. To address these challenges, we propose a solution that ensures domain-invariant prompt learning while enhancing the expressiveness of visual features. We observe that CLIP’s vision encoder struggles to identify contextual image information, particularly when image patches are jumbled up. This issue is especially severe in optical remote sensing images, where land-cover classes exhibit well-defined contextual appearances. To this end, we introduce C-SAW, a method that complements CLIP with a self-supervised loss in the visual space and a novel prompt learning technique that emphasizes both visual domain and content-specific features. We keep the CLIP backbone frozen and introduce a small set of projectors for both the CLIP encoders to train C-SAW contrastively. Experimental results demonstrate the superiority of C-SAW across multiple remote sensing benchmarks and different generalization tasks.

2 October 2023
Verified
Conference Paper
Closed access
SBSamayan Bhattacharya
Samayan Bhattacharya
ABAvigyan Bhattacharya
Avigyan Bhattacharya
SSSk Shahnawaz
Sk Shahnawaz
Abdominal pain is one of the most common symptoms for a wide range of conditions in children, under the age of 16 years. Due to the limited ability of X-ray to distinguish between structures in soft tissue, physicians often rely on Computed Tomography (CT) scan to diagnose the underlying cause of abdominal pain. A CT scan exposes the patient to 70-150 times the radiation used for an X-ray. Moreover, CT scanning equipment is often not available in low-resource countries, leading to improper diagnosis and treatment. Children are more susceptible to the harmful effects of radiation than adults and might have limited language skills, based on age, and hence limited ability to describe their symptoms to the physician. In this work, we show that it is possible to use a Machine Learning (ML) model, capable of generating synthetic CT scans, from orthogonal X-ray scans, to improve the automatic prediction of abdominal anomalies. In particular, we focus on the detection of structural anomalies such as malformed organs, cysts, and appendicitis. On average, we are able to improve the performance of the prediction model by 9.75%, with respect to the model trained on only X-ray and 4.55%, with respect to the model trained on only generated CT scan, by training it on both the generated CT scan and X-ray.

21 May 2024
Verified
Conference Paper
Open access
ABAvigyan Bhattacharya
Avigyan Bhattacharya
TCTapabrata Chakrabarti
Tapabrata Chakrabarti
SBSubhadip Basu
Subhadip Basu
AKAlistair Knott
Alistair Knott
DPDino Pedreschi
Dino Pedreschi
PBPrzemyslaw Biecek
Przemyslaw Biecek
Proliferation of online political hate speech through social media has been a persisting problem and is being recently compounded by the arrival of AI-boosted content. This can lead to wanton dissemination of misinformation/disinformation and can cause extremist radicalisation or influence national electoral processes. Given the high stakes of negative social impact, it is becoming increasingly important to address the sensitive topic of content moderation on social media platforms, the debate being the dichotomy of free speech versus content harm. From that perspective, it is crucial to establish a nuanced definition and categorisation of harmful content that is sensitive to the culture and language of the place of dissemination, which is different from the current one-size-fits-all approach, where content moderation is performed by social media companies behind closed doors. In this paper, we present a democratized solution to this problem through a crowdsourced annotation process that may be used to have a transparent method of identifying harmful content, which can then be used to make moderation decisions like contextually weighted downranking of harmful content. We present proof of concept case studies in the Indian political electoral discourse. We introduce a curated dataset of tweets labeled by multiple annotators from diverse backgrounds and visualize insightful statistical patterns emerging therefrom. This is the first stage of a multi-year Global Partnership on AI (GPAI) project on responsible AI for social media governance. In 2024 and beyond, we plan to expand the work to include both memes and tweets, that are multilingual (a mixture of Hindi/Bengali, English, and romanised Hindi/Bengali).

30 June 2025 in Neural Computing and Applications
Verified
Journal Article
Open access
PDF file
ABAvigyan Bhattacharya
Avigyan Bhattacharya
SBSubhadip Basu
Subhadip Basu
TCTapabrata Chakraborti
Tapabrata Chakraborti
Scene text image super-resolution (STISR), often considered a preliminary step for scene text recognition, refers to the task of enhancing the resolution of text embedded in natural scene images and plays a vital role in various applications. Most of the existing STISR methods either leverage deep convolutional neural networks by regarding text images as natural scene images or use a text recognizer’s feedback as guidance to the STISR process. However, since the text recognition is initially done on low-resolution images, it is mostly inaccurate, more so as the length of the words increases, thus degrading the super-resolution process. In this paper, we introduce DEPP which utilizes dictionary embedding (DE) based probabilistic priors calculated from a large English text corpus consisting of both alphabets and digits. The initial state and the bigram probabilities obtained are fused with the probability obtained from the recognizer, before passing it onto a single image super-resolution (SISR) block. By integrating DE as a prior and implementing a modified perceptual loss, the method effectively captures the contextual information of text, enabling more accurate super-resolution and visually pleasing results. Experimental results on the benchmark TextZoom dataset demonstrate that our DEPP framework achieves superior performance compared to most existing approaches, particularly for medium and long-length words, as measured by text recognition accuracy. Since DEPP uses the text recognition attributes to rectify or guide the super-resolution process, it makes our method more domain-inspired and task-aware, compared to usual black box deep learners.