The "INDIAai" (National AI Portal of India) portal provides weekly comprehensive articles highlighting the research contributions made by universities and colleges in India.

We aim to offer thorough reporting on the AI research contributions made by a single institution every week. This series allows researchers and students to provide concise explanations of their work.

Panjab University (PU) is one of India's oldest universities, founded in Lahore in 1882. It has a long history of pursuing excellence in science and technology, humanities, social sciences, performing arts, and sports education and research. At the main campus in Chandigarh, the university has 78 teaching and research departments and ten teaching and research centres/chairs. It has 188 affiliated institutions distributed over eight Punjab districts and Chandigarh's union territory, with Regional Centres in Muktsar, Ludhiana, and Hoshiarpur. It is one of the top universities in India.

Let us explore the top AI research contributions from Panjab University, Chandigarh.

Multidimensional empirical analysis of overlapping community detection methods in social networks

The authors Monika Saini and Veenu Mangat state that "The extensive scope of social network analysis gives rise to a multitude of research obstacles. Community detection represents a primary obstacle in the field of research. The literature presents a range of community detection methods, each of which has undergone an evaluation of its efficacy using a distinct set of metrics." 

However, this empirical analysis criterion for predicting the performance of a specific community detection method requires additional research and refinement. Without a multivariate framework for representing the results, earlier surveys on empirical analysis of overlapping community detection methods encountered a significant obstacle. The authors said, "The preponderance of analyses in the literature have been conducted using performance metrics exclusively. In contrast to other empirical analyses discussed in the literature, this paper examines interdependencies between different fitness metrics during community detection. Also presented is a co-performance analysis of overlapping community detection methods based on partition comparison." 

Furthermore they said that "the assessment was conducted using authentic and benchmark datasets. Researchers can use this article to guide when selecting a specific algorithm for detecting overlapping communities, based on examining partition comparison and interdependencies among fitness metrics."

SiamNet: Exploiting source camera noise discrepancies using Siamese Network for Deepfake Detection

The authors Staffy Kingra, Naveen Aggarwal and Nirmal Kaur state that "Deep neural network developments in recent times, most notably GAN (Generative Adversarial Network), have enabled the production of deepfake media that are more lifelike. This technology can alter facial expressions or swap the face of the source individual in an image or video; media that has been altered in this manner is known as "deepfake." This form of manipulated media poses Potential risks to journalism, politics, court proceedings, and numerous social aspects. Current methodologies predominantly rely on deep neural networks to extract facial artefacts directly for deepfake detection. However, these approaches fail to scrutinize nuanced inconsistencies within or between frames." 

Furthermore, contemporary deepfake detection networks exhibit improved complexity and are susceptible to overfitting particular artefacts, constraining their ability to generalize to unobserved data. This study developed a new method that uses the inconsistent noise pattern between the face region and the rest of the frame to detect altered faces in movies and photos. 

The authors said, " To compare the noise patterns of these two regions, we propose SiamNet, a two-stream Siamese-like network. By employing distinct streams, this network can extract the noise patterns from the face region and patch, thus augmenting its efficiency and effectiveness without requiring the number of parameters. Each branch comprises a pre-trained Inception-v3 architecture designed to extract camera noise."

They further said, "By employing Siamese training, comparing noise patterns generated by distinct base models is possible. It has been determined that the proposed two-branch network, SiamNet, is effective for several large-scale deepfake datasets, including FF++, Celeb-DF, DFD, and DFDC, with respective accuracy rates of 99.7%, 98.3%, 96.08%, and 89.2%." 

Improved RPPG Method to Detect BPM from Human Facial Videos

The authors Manpreet Kaur & Naveen Aggarwal state, "The COVID-19 pandemic has precipitated an upsurge in mobile applications that enable video conferencing between a physician and patient remotely. The physical parameters of patients must be assessed throughout the interaction. One of the methodologies that can be utilized to quantify heart rate in beats per minute from live facial video of patients is remote photoplethysmography (rPPG). Among the earliest supervised learning implementations of rPPG, the PhysNet model was developed."

Furthermore the authors say that "face resolution may differ depending on the camera's position. An enhanced rPPG method for detecting pulses per minute in human facial videos is presented in this work. The model undergoes validation using the UBFC dataset, to which random noise ranging from 10 to 30 per cent is added. Additionally, the dataset is expanded through the modification of image resolution. In addition, the testing outcomes on the proposed dataset were superior to those on the UBFC dataset. The experimental findings are encouraging, as evidenced by the significant reduction of 60% in RMSE and MAE values for both models."

Autoencoder-based dense denoiser and block-based wiener filter for noise reduction of optical coherence tomography

Mamta Juneja, Gurunameh Singh Chhatwal, Shatabarto Bhattacharya, Niharika Thakur, and Prashant Jindal state that "OCT, or Optical Coherence Tomography, is an advanced imaging technique that is employed to diagnose abnormalities within the retina. OCT is a technique that utilizes low-coherence light waves, specifically infrared waves, with a resolution measured in micrometres. Its purpose is to image the retinal layers within the eye." 

They say, "Using OCT to analyze the variation in thickness between distinct retinal layers can be used for diagnostic purposes. Nevertheless, diverse degrees of speckle noise hinder these strata's visibility, compromising the effectiveness of subsequent diagnostic procedures. Despite various methodologies for denoising OCT images, excessive smoothing of the images results in the obscurement of structural edge details, which consequently compromises the accuracy of the diagnosis."

They added, "Therefore, a method that effectively eliminates speckle noise while preserving crucial image details is favoured. This article describes a methodology for mitigating speckle noise in OCT images through the utilization of a Block-based Wiener Filter (BBWF) and an Autoencoder-based Dense Denoiser (ADD) neural network."