http://dspace.dtu.ac.in:8080/jspui/handle/123456789/99 2026-10-17T16:52:11Z 2026-10-17T16:52:11Z SINGH, SRISHTI Sisaudia, Varsha (SUPERVISOR) http://dspace.dtu.ac.in:8080/jspui/handle/repository/23392 2026-10-16T06:35:15Z 2026-05-01T00:00:00Z

Title: ATTENTION-GUIDED DUAL-DOMAIN EMBEDDING FOR INTELLECTUAL PROPERTY PROTECTION OF IMAGE PROCESSING MODEL Authors: SINGH, SRISHTI; Sisaudia, Varsha (SUPERVISOR) Abstract: Deep Learning based image processing models such as denoising, super-resolution and deraining consume significant computational resources, curated data and engineering effort. However when models such as those, are available through publicly accessible APIs, it renders them vulnerable to Surrogacy Model Extraction attacks, in which an adversary sends queries to the model, collects the output and training a functionally equivalent replica of the model without acquiring its parameters. Watermarking schemes typically fail to protect the model as, its reconstruction objective, is trained on the premise that the pixel level watermark embedded within it is random noise, and averaging it will result in elimination of such information. This thesis proposes an Attention Guided Dual Domain Embedding (AGDDE) network framework for the IP protection of image processing models, in the form of Frequency Aware Structure Consistent Watermarking (FA-SCW) system. A dual branch attention mechanism which is much simpler and replacing the fixed Sobel detector used in previous work, learning a differentiable mask identifying both edges and flat regions for embedding. A dual domain embedding network which combines a correct U-shape Network in spatial domain and CNN in middle frequency DCT coefficients domain. A self-supervised frequency consistency loss can be achieved to maintain spectrum invariance under geometric perturbation providing robustness to crop, rotation, JEPG compression without knowing the surrogate model. Evaluation has been performed on Color Berkeley Segmentation Dataset of 68 images (CBSD68) and DIVerse 2K resolution dataset (DIV2K), to show the imperceptibility is more than 35 dB PSNR, surrogate extraction Success Rate (SR) is higher than 85% and Bit Error Rate (BER) is less than 0.10 which is 22.9% higher than nearest competitor at higher quality. Complete system takes 30 minutes to converge on single GPU.

2026-05-01T00:00:00Z RAYKARMAKAR, KINJAL Vishwakarma, Dinesh K. (SUPERVISOR) http://dspace.dtu.ac.in:8080/jspui/handle/repository/23376 2026-10-16T06:33:14Z 2026-05-01T00:00:00Z

Title: A DUAL-STREAM CROSS-ATTENTION FRAMEWORK FOR FINE-GRAINED YOGA POSE CLASSIFICATION USING SPATIOTEMPORAL VISUAL CUES AND 3D SKELETAL GEOMETRY Authors: RAYKARMAKAR, KINJAL; Vishwakarma, Dinesh K. (SUPERVISOR) Abstract: The accurate automated recognition of yoga postures from a single video sequence presents itself as an intricate problem, lying in the intersection between computer vision and biomechanics, and holds enormous potential towards achieving superior fitness training, physical therapy, and man-machine interaction applications. Existing techniques are mainly based on either appearance features, susceptible to changes in clothing and background, or geometric keypoint features that lack contextual information. The main focus of this thesis is the proposition that a combination of the two modalities can outperform each of them independently. We introduce a complete pipeline starting from the pre-processing of untrimmed online videos of the 3DYoga90 dataset and extracting the 3D world-coordinate skeletons with the use of MediaPipe. Two novel backbones are fine-tuned on an extremely strict cross-subject split as the unimodal baselines – the first one is Channel-wise Topology Refinement Graph Convolutional Network (CTR-GCN) on the skeleton modality, and the other is Video Swin Transformer. We propose a logit-level cross-modal fusion model, termed LogitPerceiver, which merges the predictions without requiring full-fledged end to-end training due to its high computational cost. In addition, the model projects the backbone’s logits into a unified space and employs a trainable query to concentrate on two input streams as well as a gated adjustment to the skeleton prediction results. The late-fusion framework enables adaptive and context-dependent fusion of information related to the input samples from both visual and geometric modalities. Experiments carried out using the 3DYoga90 dataset reveal that the skeleton-based CTR GCN can achieve a balanced accuracy of 86.96% compared to 81.64% by the Video Swin Transformer. The proposed fusion framework is able to score 89.64% balanced accuracy, 91.06% overall accuracy, and 88.96% macro-averaged F1-score, beating the performance of all unimodal baselines and establishing a new baseline on the dataset. This thesis provides a reproducible multimodal framework, a light fusion network compatible with frozen experts, and a strong benchmark dataset for yoga position classification.

2026-05-01T00:00:00Z SARKAR, ANWESH Vishwakarma, Dinesh K. (SUPERVISOR) http://dspace.dtu.ac.in:8080/jspui/handle/repository/23375 2026-10-16T06:33:08Z 2026-05-01T00:00:00Z

Title: DEVELOPMENT OF FRAMEWORK FOR 2D TO 3D RECONSTRUCTION AND VOLUMETRIC SEGMENTATION Authors: SARKAR, ANWESH; Vishwakarma, Dinesh K. (SUPERVISOR) Abstract: There The ever-growing requirements for swift content generation in gaming, e commerce, and VR applications prompt the necessity of fast reconstruction from single view. In this thesis, we present an enhanced pipeline for generating textured 3D models from one image through the integration of deep image segmentation with feed-forward volumetric inference. For background removal, we propose a U²‑Net-based segmentation module that outputs a normalized foreground mask which then goes through TripoSR - a vision transformer encoder paired with a lightweight triplane architecture that predicts an occupancy map. We obtain a watertight mesh using marching cubes. The quality of our reconstruction, measured on a synthetic spherical object, is confirmed through the following metrics: CD of 1.857 × 10⁻³, volumetric IoU of 0.7786, and HD of 0.0849, providing excellent shape recovery performance. The presented method significantly outperforms TripoSR without image segmentation (CD = 10.2 × 10⁻³) and the OccNet [12] model (CD = 12.4 × 10⁻³), achieving a geometric error reduction by almost five times. The whole procedure takes less than two seconds when run on a consumer-grade GPU. Our ablation studies show that the introduction of a separate 2D segmentation stage is crucial for obtaining high-fidelity results.

2026-05-01T00:00:00Z SAINI, SARTHAK Sisaudia, Varsha (SUPERVISOR) http://dspace.dtu.ac.in:8080/jspui/handle/repository/23374 2026-10-16T06:33:01Z 2026-05-01T00:00:00Z

Title: A Lightweight ATTENTION- GUIDED BIFPN YOLO11 model FOR DRONE DETECTION Authors: SAINI, SARTHAK; Sisaudia, Varsha (SUPERVISOR) Abstract: With the fast increase in the number of applications that UAVs perform in areas of surveillance, security, disaster management, and commercial operations, there has developed a very significant requirement for efficient and accurate real-time drone detection systems. Nevertheless, the process of detecting UAVs poses great challenges owing to factors such as small object size in image frames, changes in illumination levels, complexity of scenes and motion variations. While YOLO-based deep learning methods achieve real-time capabilities, they still struggle in detecting small UAVs and coping with complex environments. This thesis presents a state-of-the-art UAV object detector model based on YOLO11. This architecture is enhanced using two additional models – the Convolutional Block Attention Module (CBAM) and the Bi - directional Feature Pyramid Network (BiFPN). The purpose is to increase the ability of this model to detect smaller objects while maintaining efficiency and fast operation. The first enhancement allows improving the detection of key features of the drone objects both spatially and through channel attention. The other modification makes it possible for the model to improve its multi-level feature representations since it facilitates two-way fusion of the features from different pyramid levels, which helps in localization of the drones at different levels. The models were trained and tested using the open - source dataset YOLO Drone Detection Dataset, available at Kaggle, that contains numerous examples of indoor, outdoor, distant, and outline drones having 347 samples for validation along with 369 labeled drones. Both YOLO11n and YOLO11s models were experimented with extensively using the same training set-up. For performance evaluation, the standard object detection parameters such as Precision , Recall , mAP@50 , mAP@50-95 , F1-Score and Frames Per Second (FPS) were used. The experimental results show that YOLO11s + CBAM + BiFPN obtained a Precision of 94.8%, Recall of 93.7%, mAP@50 of 96.1%, mAP@50-95 of 71.8% and F1-Score of 94.2%. These correspond to gains of +2.4%, +5.6%, +2.5%, and +3.3% respectively compared to the plain YOLO11s baseline, while still running in real-time at 75 FPS on an NVIDIA Tesla T4 GPU. Further analysis of qualitative confidence demonstrated that YOLO11s + CBAM + BiFPN had an average detection confidence of 16.4% higher than YOLO11n + CBAM + BiFPN on the same input frames, indicating the benefit of deeper feature extraction for small drone targets. v Collectively, the results indicate that the combination of lightweight attention modules and advanced multi - scale features and fusion can significantly improve the accuracy and robustness of UAV detection. The devised approach exhibited strong robustness under difficult circumstances, including extremely small and faraway drones, reflection, occlusions, and contrastless images. This is a viable and scalable solution to many practical problems such as drone tracking and surveillance.

2026-05-01T00:00:00Z