Drones are increasingly used in civilian and defense domains, but reliable detection remains challenging due to their small size, fast motion, and diverse environments. Existing datasets, such as synthetic benchmarks, fail to capture real-world variability. We introduce RealDroneVision, a unified contribution that advances both dataset and methodology. First, we curate a large-scale real-world drone detection dataset comprising 173,023 images, constructed via a semi-automatic pipeline inspired by self-annotated labeling from videos, enhanced with a human-in-the-loop to iteratively reduce false positives and false negatives. This approach yields high-quality annotations with reduced manual effort. Second, we propose the Nano Object Vision Attention (NOVA) module, a drop-in replacement for YOLOv8’s C2f block. By combining depthwise separable convolutions, scale-aware dilated branches, lightweight mixing, and coordinate-aware attention, our design improves small-object detection while remaining computationally efficient. Extensive benchmarks against YOLOv8m/l and YOLOv9c/e demonstrate that YOLOv8-NOVA dominates across precision (0.912), recall (0.870), and mAP@50 (0.920) while being significantly more lightweight (2.3M params, 5 MB weights). These results establish RealDroneVision as a strong foundation for advancing real-world drone detection research.

Data augmentation has been a prevalent approach in improving the performance of deep learning models against slight variations in data. Adversarial learning is one such form of data augmentation. In this work, we aim to introduce a framework to generate harder examples for a specific object class and an adversarial attack for the object detection task. We have also presented our study on the effect of training against such generated harder examples and adversarial samples in object detection. We have applied this adversarial learning technique to a YOLOv3 model and due to the nature of the attack, we demonstrated a substantial improvement in average precision (AP) for a single class of the COCO dataset. As per the literature, we are the first to introduce this kind of class-specific data augmentation strategy in object detection. With our approach, we have shown an improvement of 23.34% in AP for Cat class and 3.1% on overall mAP of YOLOv3 model on clean validation data, while 43.5% improvement in AP for the Cat class on the composite images with class-specific adversarial samples.

Phase unwrapping is a challenging task in signal processing, spanning its applications in optical metrology, SAR interferometry, and many other signal reconstruction tasks. Fringe Projection Profilometry is a popular active-sensing approach for generating high-resolution three-dimensional (3D) surface information in which phase unwrapping is a crucial step. This letter proposes a multi-task learning-based phase unwrapping method for simultaneous denoising and wrap-count prediction in fringe projection. The proposed network, referred to as TriNet, has nested pyramidal architecture with a single encoder and two decoders, all connected through skip connections. The proposed approach does not require any pre-processing for noise removal like the conventional methods or any post-processing such as smoothing, like in existing deep learning methods but results in a quite accurate phase unwrapping. The proposed method outperforms the existing and state-of-the-art methods for the 3D reconstruction task in Fringe Projection by a significant margin even in the presence of very high noise.

High Dynamic Range (HDR) deghosting is an indispensable tool in capturing wide dynamic range scenes without ghosting artifacts. Recently, convolutional neural networks (CNNs) have shown tremendous success in HDR deghosting. However, CNN-based HDR deghosting methods require collecting large datasets with ground truth, which is a tedious and time-consuming process. This paper proposes a pioneering work by introducing zero and few-shot learning strategies for data-efficient HDR deghosting. Our approach consists of two stages of training. In stage one, we train the model with few labeled (5 or less) dynamic samples and a pool of unlabeled samples with a self-supervised loss. We use the trained model to predict HDRs for the unlabeled samples. To derive data for the next stage of training, we propose a novel method for generating corresponding dynamic inputs from the predicted HDRs of unlabeled data. The generated artificial dynamic inputs and predicted HDRs are used as paired labeled data. In stage two, we finetune the model with the original few labeled data and artificially generated labeled data. Our few-shot approach outperforms many fully-supervised methods in two publicly available datasets, using as little as five labeled dynamic samples.

The blood smear analysis provides vital information and forms the basis to diagnose most of the diseases. With recent developments, deep learning methods can analyze the microscopic blood sample using image processing and classification tasks with less human effort and increased accuracy. In this work, embarking upon domain-specific feature extraction and active learning, we propose a compact, yet efficient feature fusion ensemble-based architecture for WBC sub-class classification and WBC disease identification which can automate and speed up the process of blood smear analysis with the help of digital image slide scanners.The proposed architecture is a three-stage multi-channel architecture with shallow feature inputs, deep feature extractors, and classification stages, respectively. The trainable parameters are quite less in our architecture when compared to deeper networks like ResNet 152, VGG19, which are the State of the Art (SOTA) methods [1], [2], [3]. However, labeling large medical data sets has been challenging and very costly. To mitigate huge labeling requirements and costs, Active Learning is employed to train this architecture and demonstrated much higher accuracy with quite less labeled data than SOTA. The proposed approach is shown to be quite general and yields better performance in terms of accuracy in WBC classification and Disease identification as well, with much fewer labeled samples for training, when compared with recent approaches employing deeper models.

Recently deep learning-based Siamese networks with region proposals for visual object tracking are becoming popular. These frameworks, while testing, perform extra computations on the output of a trained network to predict the bounding box (bbox). This process hinders end-to-end training of the above class of networks and hampers the precise estimation of the bbox in testing. In this paper, we propose a framework close to the Siamese class of networks, but guided by Intersection Over Union (IOU) to predict precise bbox directly in the image space rather than at the feature space. To maximise the IOU of predicted bbox with respect to ground truth, we introduce a new module and corresponding loss function in training the network. The proposed approach enables end-to-end training and testing under similar lines, circumventing the typical bottleneck of the existing Siamese trackers. When evaluated on VOT2018 and GOT-10k tracking benchmarks, the proposed approach outperformed the base approach by more than 10% in terms of average overlap and compares favourably to state-of-the-art methods.

Phase reconstruction in Digital Holographic Interferometry (DHI) is widely employed for 3D deformation measurements of the object surfaces. The key challenge in phase reconstruction in DHI is in the estimation of the absolute phase from noisy reconstructed interference fringes. In this paper, we propose a novel efficient deep learning approach for the phase estimation from noisy interference fringes in DHI. The proposed approach takes noisy reconstructed interference fringes as input and estimates the 3D deformation field or the object surface profile as the output. The 3D deformation field measurement of the object is posed as the absolute phase estimation from the noisy wrapped phase, that can be obtained from the reconstructed interference fringes through arctan function. The proposed deep neural network is trained to predict the fringe-order through a fully convolutional semantic segmentation network, from the noisy wrapped phase. These predictions are improved by simultaneously minimizing the regression error between the true phase corresponding to the object deformation field and the estimated absolute phase considering the predicted fringe order. We compare our method with conventional methods as well as with the recent state-of-the-art deep learning phase unwrapping methods. The proposed method outperforms conventional approaches by a large margin, while we can observe significant improvement even with respect to recently proposed deep learning-based phase unwrapping methods, in the presence of noise as high as 0dB to -5dB.

Phase unwrapping is an ill-posed classical problem in many practical applications of significance such as 3D profiling through fringe projection, synthetic aperture radar and magnetic resonance imaging. Conventional phase unwrapping techniques estimate the phase either by integrating through the confined path (referred to as path-following methods) or by minimizing the energy function between the wrapped phase and the approximated true phase (referred to as minimum-norm approaches). However, these conventional methods have some critical challenges like error accumulation and high computational time and often fail under low SNR conditions. To address these problems, this paper proposes a novel deep learning framework for unwrapping the phase and is referred to as “PhaseNet 2.0”. The phase unwrapping problem is formulated as a dense classification problem and a fully convolutional DenseNet based neural network is trained to predict the wrap-count at each pixel from the wrapped phase maps. To train this network, we simulate arbitrary shapes and propose new loss function that integrates the residues by minimizing the difference of gradients and also uses L1loss to overcome class imbalance problem. The proposed method, unlike our previous approach PhaseNet, does not require post-processing, highly robust to noise, accurately unwraps the phase even at the severe noise level of -5 dB, and can unwrap the phase maps even at relatively high dynamic ranges. Simulation results from the proposed framework are compared with different classes of existing phase unwrapping methods for varying SNR values and discontinuity, and these evaluations demonstrate the advantages of the proposed framework. We also demonstrate the generality of the proposed method on 3D reconstruction of synthetic CAD models that have diverse structures and finer geometric variations. Finally, the proposed method is applied to real-data for 3D profiling of objects using fringe projection technique and digital holographic interferometry. The proposed framework achieves significant improvements over existing methods while being highly efficient with interactive frame-rates on modern GPUs.

Due to high variability in shape, structure and occurrence; the non-palpable breast masses are often missed by the experienced radiologists. To aid them with more accurate identification, computer-aided detection (CAD) systems are widely used. Most of the developed CAD systems use complex handcrafted features which introduce difficulties for further improvement in performance. Deep or high-level features extracted using deep learning models already have proven its superiority over the low or middle-level handcrafted features. In this paper, we propose an automated deep CAD system performing both the functions: mass detection and classification. Our proposed framework is composed of three cascaded structures: suspicious region identification, mass/no-mass detection and mass classification. To detect the suspicious regions in a breast mammogram, we have used a deep hierarchical mass prediction network. Then we take a decision on whether the predicted lesions contain any abnormal masses using CNN high-level features from the augmented intensity and wavelet features. Afterwards, the mass classification is carried out only for abnormal cases with the same CNN structure. The whole process of breast mass classification including the extraction of wavelet features is automated in this work. We have tested our proposed model on widely used DDSM and INbreast databases in which mass prediction network has achieved the sensitivity of 0.94 and 0.96 followed by a mass/no-mass detection with the area under the curve (AUC) of 0.9976 and 0.9922 respectively on receiver operating characteristic (ROC) curve. Finally, the classification network has obtained an accuracy of 98.05% in DDSM and 98.14% in INbreast database which we believe is the best reported so far.

Online Signature Verification (OSV) is a widely used biometric feature for recognized and authorized technique to authenticate a writers's distinctiveness and behavioral characteristic. Owing to huge intra-individual changeability, OSV is a challenging problem. Usage of online signatures in m-commerce and m-payment demands for light weight frameworks to classify a signature. The recent OSV models grounded on convolutional neural networks (CNN) and its variants are heavy weight and computationally intensive due to higher amount of parameters to learn. In this context, we put forward a CNN centric OSV framework which uses a stack of depthwise separable (DWS) convolution layers, which makes the framework light weight and enables the few shot learning for signature verification with quite higher accuracy compared to conventional deep learning models. To prove the robustness of our proposed framework, we performed exhaustive experimental evaluations with three standard datasets i.e. MCYT-100 (DB1), SUSIG-Visual corpus and SVC-2004-Task2. Experimental results confirm the efficiency of depthwise separable convolutions grounded OSV by realizing a lesser error rate as related to various current and state-of-the art OSV frameworks.

Phase unwrapping is a crucial signal processing problem in several applications that aims to restore original phase from the wrapped phase. In this letter, we propose a novel framework for unwrapping the phase using deep fully convolutional neural network termed as PhaseNet. We reformulate the problem definition of directly obtaining continuous original phase as obtaining the wrap-count (integer jump of 2 π) at each pixel by semantic segmentation and this is accomplished through a suitable deep learning framework. The proposed architecture consists of an encoder network, a corresponding decoder network followed by a pixel-wise classification layer. The relationship between the absolute phase and the wrap-count is leveraged in generating abundant simulated data of several random shapes. This deliberates the network on learning continuity in wrapped phase maps rather than specific patterns in the training data. We compare the proposed framework with the widely adapted quality-guided phase unwrapping algorithm and also with the well-known MATLAB's unwrap function for varying noise levels. The proposed framework is found to be robust to noise and computationally fast. The results obtained highlight that deep convolutional neural network can indeed be effectively applied for phase unwrapping, and the proposed framework will hopefully pave the way for the development of a new set of deep learning based phase unwrapping methods.

Atmospheric motion vectors (AMVs) estimation helps in better understanding of atmospheric dynamics and also plays a key role in weather forecasting. It has been a challenging task because of the nonrigid motion of clouds and cyclones. In this paper, a modified Weighted Ensemble Transform Kalman Filter-based data assimilation technique is proposed for accurate flow vector estimation at each pixel directly from satellite generated infrared images of clouds/cyclones. This method provides clear visualization of both local and global motion with spatial and temporal consistencies very efficiently even in the case of splitting and merging of clouds or over long tracks. One of the key abilities of proposed method is in forecasting applications and also for generating motion vectors in the absence of data in real scenarios, even without the usage of the existing complex weather models. Estimated AMVs are validated using state-of-the-art European Centre for Medium-range Weather Forecasting (ECMWF) analysis data, and cyclone tracks are validated using the Indian Meteorological Department (IMD) best track data. The results obtained demonstrate the efficacy of proposed method over other existing methods.

Visual Object tracking research has undergone significant improvement in the past few years. The emergence of tracking by detection approach in tracking paradigm has been quite successful in many ways. Recently, deep convolutional neural networks have been extensively used in most successful trackers. Yet, the standard approach has been based on correlation or feature selection with minimal consideration given to motion consistency. Thus, there is still a need to capture various physical constraints through motion consistency which will improve accuracy, robustness and more importantly rotation adaptiveness. Therefore, one of the major aspects of this paper is to investigate the outcome of rotation adaptiveness in visual object tracking. Among other key contributions, the paper also includes various consistencies that turn out to be extremely effective in numerous challenging sequences than the current state-of-the-art.

In this paper, we introduce an active learning (AL)-based object training library generation for a multiclassifier object-oriented image analysis (OOIA) system. While several AL approaches do exist for pixel-based training library generation and for hyperspectral image classification, there is no standard training library generation strategy for OOIA of very high spatial resolution images. Given a sufficient number of training samples, supervised classification is the method of choice for image classification. However, this strategy becomes computationally expensive with the increase in the number of classes or the number of images to be classified. The above-mentioned issue is solved in this proposed method, where an optimized training library of objects (superpixels) is generated based on a batch mode AL approach. A softmax classifier is used as a detector in this method, which helps in determining the right samples to be chosen for library updation. To this end, we construct a multiclassifier system with max-voting decision to classify an image at pixel level. This algorithm was applied on three different very high-resolution airborne data sets, each with varying complexity in terms of variations in geographical context, sensors, illumination, and view angles. Our method has empirically outperformed the traditional OOIA by producing equivalent accuracy with a training library that is orders of magnitude smaller. In addition, the most distinctive ability of the algorithm is experienced in the most heterogeneous data set, where its performance in terms of accuracy is around twice the performance of the traditional method in the same situation. The generality of this classification strategy is proved through its performance on multispectral images and for cross-domain application. Finally, the robustness of this method is identified by comparing its performance with an alternative AL approach-self-learning-based semisupervised SVM. The capability of the proposed method to handle highly heterogeneous data is identified as the primary reason for its robustness.

Although visual object tracking algorithms are capable of handling various challenging scenarios individually, none of them are robust enough to handle all the challenges simultaneously. For any online tracking by detection method, the key issue lies in detecting the target over the whole frame and updating systematically a target model based on the last detected appearance to avoid the drift phenomenon. This paper aims at proposing a novel robust tracking algorithm by fusing the frame level detection strategy of tracking, learning, & detection with the systematic model update strategy of Kernelized Correlation Filter tracker. The risk of drift is mitigated by the fact that the model updates are primarily driven by the detections that occur in the spatial neighborhood of the latest detections. The motivation behind the selection of trackers is their complementary nature in handling tracking challenges. The proposed algorithm efficiently combines the two state-of-the-art tracking algorithms based on conservative correspondence measure with strategic model updates, which takes advantages of both and outperforms them on their short ends by virtue of other. Extensive evaluation of the proposed method based on different metrics is carried out on the data sets ALOV300++, Visual Tracker Benchmark, and Visual Object Tracking. We demonstrated its performance in terms of robustness and success rate by comparing with state-of-the-art trackers.

This paper presents a novel, efficient scheme for the analysis of Sea Surface Temperature (SST) ocean images. We consider the estimation of the velocity fields and vorticity values from a sequence of oceanic images. The contribution of this paper lies in proposing a novel, robust and simple approach based on Weighted Ensemble Transform Kalman filter (WETKF) data assimilation technique for the analysis of real SST images, that may contain coast regions or large areas of missing data due to the cloud cover.

We propose a recursive spatial-domain speckle reduction algorithm for synthetic aperture radar (SAR) imagery based on the unscented Kalman filter (UKF) with a discontinuity-adaptive Markov random field (DAMRF) prior. The capability of the UKF in handling speckle noise and the feature preservation ability of the DAMRF model are explored within a unified framework through importance sampling.

This correspondence proposes a recursive algorithm for noise reduction in synthetic aperture radar imagery. Excellent despeckling in conjunction with feature preservation is achieved by incorporating a discontinuity-adaptive Markov random field prior within the unscented Kalman filter framework through importance sampling. The performance of this method is demonstrated on both synthetic and real examples.

This paper presents a novel approach based on the unscented Kalman filter (UKF) for image estimation in film-grain noise. The image prior is modeled as non-Gaussian and is incorporated within the UKF frame work using importance sampling. A small carefully chosen deterministic set of sigma points is used to capture the prior and is propagated through film-grain nonlinearity to compute image statistics. Experimental results are given to demonstrate the efficacy of the proposed method.

This paper presents discontinuity adaptive image estimation within the Kalman filter framework by non-Gaussian modeling of the image prior. A generalized methodology is proposed for specifying state-dynamics using the conditional density of the state given its neighbors, without explicitly defining the state equation. The novelty of our approach lies in directly obtaining the predicted mean and variance of the non-Gaussian state conditional density by importance sampling and incorporating them in the update step of the Kalman filter. Experimental results are given to demonstrate the effectiveness of the proposed method in preserving edges.