Detecting camouflaged objects in agriculture is a critical computer vision challenge, particularly for green fruits blending with dense foliage. This study introduces DLP-14k, a novel dataset of 14k high-resolution images capturing lime fruits under diverse real-world conditions, including varying lighting, occlusions, and negative samples (e.g., foliage-only images). Unlike general-purpose datasets like COCO or agricultural datasets like PlantDoc, DLP-14k uniquely focuses on camouflaged fruit detection, enhanced by data augmentation (rotation, flipping, brightness adjustment) to mitigate its small size. We evaluate YOLOv8 and Faster R-CNN for object detection and segmentation, showing YOLOv8’s efficiency for real-time applications but higher false positives in low-contrast settings, while Faster R-CNN achieves superior accuracy in occluded environments at greater computational cost. By introducing DLP-14k and analyzing these trade-offs, this study offers insights for precision agriculture. Future work will expand the dataset and explore hybrid models, advancing applications in complex visual environments.

Detecting camouflaged objects in agricultural environments-particularly green lime fruits obscured by dense foliage-remains challenging due to low color and texture contrast. This paper introduces DLP-14k, a benchmark dataset of 14,000 high-resolution field images encompassing varied illumination conditions, multiple occlusion levels, and 1,000 foliage-only negative samples. The dataset is augmented using controlled rotations, flips, and contrast adjustments to enhance generalization. We evaluate ten state-of-the-art object detection frameworks, including YOLOv8, Faster R-CNN, EfficientDet, RetinaNet, and MobileNet-SSD, covering both singleand two-stage architectures. Experimental results show that Faster R-CNN attains the highest$\mathbf{m A P}$(0.97) and accuracy (0.93), YOLOv8 balances precision and real-time performance (>45 FPS), and MobileNet-SSD achieves the fastest inference (45.2 FPS) with moderate accuracy. Ablation studies on augmentation and backbone depth highlight their effects on mAP and recall. The dataset, baseline code, and standardized protocols are released to enable reproducible research and to support advanced vision systems for autonomous harvesting and field monitoring.

Radio Frequency Interference (RFI) from cellular and other communication networks is commonly mitigated at the radio telescope without any active collaboration with the interfering sources. The expanding Universe and simultaneous proliferation of Earth-based and LEO communication infrastructure is causing unprecedented RFI that require collaborative strategies to maintain the scientific and societal goals of each. In this work, we develop deep learning based models that enable collaboration with minimal overhead while also providing accurate RFI characterization and simplified cancellation strategies. This multistage system design is adaptable to changing statistics of the RFI signals generated from cellular networks and allows single step RFI cancellation by signal processing chain modeling (e.g. filtering and digitization loss) at the Telescope. Through our analysis and simulation using real astronomical signals, we are able to remove RFI generated from cellular networks with comparable accuracy to the state of the art with only 25% of the communication overhead and overall reduced computation complexity from O(n3) to O(n2).

Ultra-broadband communication in emerging spectrum, like Terahertz (THz) band, is the frontier to meet the data rate requirements of future wireless communication systems. Existing signal processing based methods are developed for sub-6 GHz band, which cannot capture the intricacies in ultra-broad THz bandwidth and non-linearities arising from hardware. To overcome these limitations, we develop neural network (NN) models for OFDM receiver, where expert knowledge of wireless communication is infused in different stages and parameters of the model to create a practical receiver that can adapt to different wireless environments. The parameters of the NN are derived from underlying theory and can be adapted to different wireless environments. Our model is designed to capture the correlation between real and imaginary components of wireless signals, that can be trained with limited data. The models are trained with over-the-air captured OFDM signals, transmitted in THz band with 10 GHz bandwidth. Our results show significant improvement in bit error rate (BER) for different modulation orders (upto 6 dB in BPSK and 1.2 dB in QAM 64) compared to existing signal processing based receiver designs.

Channel estimation for Orthogonal Frequency Division Multiplexing (OFDM) transmission is well investigated with model based approaches. Recent effort also explores the data driven approaches to exploit the capabilities of Neural Networks (NNs) to estimate the channel. These models are mostly being developed as black box without any anchor to the theory of wireless signal propagation. We propose a NN model, where the structure and parameters are derived from the domain knowledge of wireless signal and channel characteristics. Our model is developed in two stages: the first stage handles the noise reduction, while the second stage extracts the channel characteristics to reduce error caused by multipath. We have used the knowledge of signal to noise ratio and subcarrier correlation due to channel delay spread to empower the two stages of the proposed model. Our results also show that induction of domain knowledge results in reduction of data dependency by 60%. Our model outperforms the practical model based methods as well as blind data driven approaches. It achieves ∼10 dB improvement over Least Square channel estimation.

In many sequential decision-making problems (e.g., robotics control, game playing, sequential prediction), human or expert data is available containing useful information about the task. However, imitation learning (IL) from a small amount of expert data can be challenging in high-dimensional environments with complex dynamics. Behavioral cloning is a simple method that is widely used due to its simplicity of implementation and stable convergence but doesn't utilize any information involving the environment’s dynamics. Many existing methods that exploit dynamics information are difficult to train in practice due to an adversarial optimization process over reward and policy approximators or biased, high variance gradient estimators. We introduce a method for dynamics-aware IL which avoids adversarial training by learning a single Q-function, implicitly representing both reward and policy. On standard benchmarks, the implicitly learned rewards show a high positive correlation with the ground-truth rewards, illustrating our method can also be used for inverse reinforcement learning (IRL). Our method, Inverse soft-Q learning (IQ-Learn) obtains state-of-the-art results in offline and online imitation learning settings, significantly outperforming existing methods both in the number of required environment interactions and scalability in high-dimensional spaces, often by more than 3x.

Learned neural solvers have successfully been used to solve combinatorial optimization and decision problems. More general counting variants of these problems, however, are still largely solved with hand-crafted solvers. To bridge this gap, we introduce belief propagation neural networks (BPNNs), a class of parameterized operators that operate on factor graphs and generalize Belief Propagation (BP). In its strictest form, a BPNN layer (BPNN-D) is a learned iterative operator that provably maintains many of the desirable properties of BP for any choice of the parameters. Empirically, we show that by training BPNN-D learns to perform the task better than the original BP: it converges 1.7x faster on Ising models while providing tighter bounds. On challenging model counting problems, BPNNs compute estimates 100's of times faster than state-of-the-art handcrafted methods, while returning an estimate of comparable quality.

Recently, near-field source localization has received more attention due to its capability of estimating the source direction of arrival (DOA) as well as the distance of the source from sensors. In the near-field, the waves are spherical and form a second-order model. A two dimensional (2D) method such as 2D MUSIC is used to solve this second-order problem. Theoretically, the 2D MUSIC has infinite resolution, but practically its resolution is limited and depends on various parameters. In this paper, we have investigated the resolution limit of 2D MUSIC. It has been seen that for closely spaced sources, the resolvability of the MUSIC depends upon the parameters such as the number of sensors, and the distance between the sensors. Mathematically this relationship is derived based on the orthogonality property of the autocorrelation matrix's eigenvectors. Simulation results varying above mentioned parameters validate the mathematical analysis. This study is important for practical implementation, where closely spaced sources are present.

Insulated gate bipolar transistors (IGBT) are used broadly in DC power transmission, power converters, and drives. These applications are cost-sensitive and require high module reliability. Researchers have been found that IGBT degradation over time is directly dependent on its junction temperature. Normally the temperature is measured with a temperature sensor close to the IGBT case or module. This requires the predesigning of the converter or results in a cost-effective and inconvenient process. In this paper, we are proposing a non-invasive method of junction temperature extraction, which do not suffer from the difficulties mentioned above. The method is based on the analysis of electromagnetic radiation (EMR) generated by the IGBT. It has been found that the radiation is a function of switching delay of the IGBT, and this delay is directly proportional to the junction temperature. Thus the junction temperature of IGBT is extracted from the EMR. The method requires only a single loop antenna to capture the radiation. Experimentally to increase the junction temperature of IGBT, an accelerated aging method is adopted. Then the junction temperature, the delay time, and the electromagnetic radiations are measured. The junction temperature is extracted from the inverse relation of these three parameters.