Self-supervised Learning (SSL) has become a powerful paradigm for representation learning without manual annotations. However, most existing frameworks focus on global alignment and struggle to capture the hierarchical, multi-scale lesion patterns characteristic of plant disease imagery. To address this gap, we propose PSMamba , a progressive self-supervised framework that integrates the efficient sequence modelling of Vision Mamba (VM) with a dual-student hierarchical distillation strategy. Unlike conventional single teacher-student designs, PSMamba employs a shared global teacher and two specialised students: one processes mid-scale views to capture lesion distributions and vein structures, while the other focuses on local views to capture fine-grained cues such as texture irregularities and early-stage lesions. This multi-granular supervision facilitates the joint learning of contextual and detailed representations, with consistency losses ensuring coherent cross-scale alignment. Experiments on three benchmark datasets show that PSMamba consistently outperforms representative CNN-, Transformer-, SSL-, and Mamba-based baselines, delivering superior accuracy and robustness in both domain-shifted and fine-grained scenarios.

Abstract This paper proposes MFDP-LeafNet, a novel framework designed to address the challenges associated with leaf classification using few-shot learning (FSL) with limited labelled data. Existing FSL models often struggle with insufficient training data, rely on additional learning strategies, suffer from limited feature diversity, and lack adaptability during inference. These issues limit the generalisation capability, particularly when dealing with variations in leaf shape, texture, and lighting conditions. To overcome these challenges, MFDP-LeafNet introduces a multi-backbone feature integration (MFI) module and a novel historical dynamic prototypical network (HDPN) module. Firstly, the MFI module leverages multiple pre-trained models to capture diverse feature representations, improving the model’s generalisation across diverse leaf categories. To address the computational overhead of this approach, a sparse autoencoder is incorporated with the MFI module, which selects the most discriminative features, thereby optimising computational efficiency. Secondly, the HDPN module enables adaptive learning during inference without the need for additional training. To validate the robustness of MFDP-LeafNet, we have conducted experiments across a diverse set of datasets, including three plant leaf datasets (Swedish Leaf, Flavia Leaf, Egyptian Plant Leaf), two widely used FSL benchmark datasets (MiniImageNet and FC100), and one plant disease dataset (PlantVillage). The experimental results across all datasets demonstrate that MFDP-LeafNet achieves competitive or improved performance compared with state-of-the-art models. The code is available at the following link: https://github.com/ismailmmu/MFDP-LeafNet .

Plant Disease Detection (PDD) is a key aspect of precision agriculture. However, existing deep learning methods often rely on extensively annotated datasets, which are time-consuming and costly to generate. Self-supervised Learning (SSL) offers a promising alternative by exploiting the abundance of unlabeled data. However, most existing SSL approaches suffer from high computational costs due to convolutional neural networks or transformer-based architectures. Additionally, they struggle to capture long-range dependencies in visual representation and rely on static loss functions that fail to align local and global features effectively. To address these challenges, we propose ConMamba, a novel SSL framework specially designed for PDD. ConMamba integrates the Vision Mamba Encoder (VME), which employs a bidirectional State Space Model (SSM) to capture long-range dependencies efficiently. Furthermore, we introduce a dual-level contrastive loss with dynamic weight adjustment to optimize local-global feature alignment. Experimental results on three benchmark datasets demonstrate that ConMamba significantly outperforms state-of-the-art methods across multiple evaluation metrics. This provides an efficient and robust solution for PDD.

This paper presents an integrated deep learning framework that unifies fine-grained corner, edge, and surface evaluation into a coherent grading pipeline, enabling consistent, scalable, and explainable trading card grading for real-world deployment. The proposed system decomposes each card into multiple high-resolution regions—eight corners (front and back), edge patches, and surface segments—each evaluated by specialized deep learning models optimized for distinct defect types. Model outputs are systematically aggregated to produce an overall grade, ensuring both granularity and consistency in evaluation. The modular design supports real-time inference and seamless deployment via FastAPI, enabling scalable industrial adoption. We detail the architecture, deployment pipeline, and performance evaluation, and share key lessons from production integration. Tested on a large industrial dataset, the system achieves high accuracy and exhibits strong agreement with expert human graders. By delivering explainable, consistent, and costefficient grading, this work provides a practical, production-ready alternative to traditional manual grading, addressing a critical bottleneck in the collectible card industry.

The rapid growth of Internet of Things (IoT) devices increases the need for robust and efficient device identification techniques to safeguard network security and integrity. Device fingerprinting emerges as a promising approach for identifying IoT devices; however, existing methods require substantial computational and memory resources, which limits their suitability for real-time and resource-constrained environments. This study presents a novel IoT device fingerprinting technique based on the frequency of communication, leveraging the inherent communication patterns of IoT devices. The proposed technique is computationally lightweight, resource-efficient, and adaptable to dynamic network environments. To enhance transparency and trust, this work incorporates explainable machine learning techniques by leveraging feature importance scores to improve the interpretability of classification decisions and reveal the contribution of communication frequency features to model predictions. Experimental evaluation using multiple benchmark datasets shows that the proposed technique has excellent performance, achieving an accuracy of up to 99.9% and demonstrating minimal inference time (as low as 0.001 seconds). It significantly outperforms existing methods by reducing computational overhead and offers a practical and scalable solution for real-world IoT deployments.

In real-world scenarios, Hyperspectral Image (HSI) datasets introduce potential noise inaccuracies due to multiple annotators. Label noise poses a significant challenge for practical deep learning, yet this issue is largely unexplored. Existing methods, which attempt to clean the noisy labelled data to increase classification accuracy, are computationally expensive and face the risk of removing correctly labelled data. In contrast, other methods that work with noisy labelled data but attempt to minimise the noise impact on classification by formulating a robust loss function lose classification accuracy when the ratio of incorrectly to correctly labelled data is high. This work proposes a Dual Spectral-Spatial Residual Adaptive (DSSRA) network to minimise the noise effect even when the amount of noisy labelled data is high. It offers the following contributions: (1) effective salient feature extraction modules to enhance the discriminatory representation of different classes in the proposed DSSRA network; (2) an adjusted noise tolerance loss (ANTL) function that down-weights the impact of learning with noisy labels. ANTL combines normalised focal loss and reverse cross-entropy to counter label noise; and (3) extensive testing on noisy versions of several benchmark HSI datasets. The results show that our DSSRA model outperforms the state-of-the-art HSI classification methods in handling noisy labels, offering a robust solution for real-world applications.

The trading card market is a dynamic industry where card values depend on meticulous grading of condition, authenticity, and quality. Traditionally, grading is performed manually by expert assessors, but this process is prone to subjectivity and inconsistency. Automated grading systems could ensure greater objectivity and uniformity in assessments. Key grading factors include centering, edges, corners, and surface quality. While our previous work focused on corner grading, this paper explores edge grading using CNN and transfer learning models such as DenseNet, ResNet, and VGG. By fine-tuning, ResNet50 achieved 93% accuracy on a dataset from our industry partner. To address uncertainty in grading, various calibration methods are employed, and a human-in-the-loop approach enhances robustness. A final scoring method provides an objective edge grading for each card.

Transmission line safety is vital to the nation’s economy and daily lives. In recent decades, most electric utilities have relied on light detection and ranging (LiDAR) technology to inspect transmission line corridors. However, occlusion and point density gaps in scanned LiDAR data will affect power line extraction. Moreover, high-voltage transmission line (HVTL) corridors in complicated locations with multiple loops pose significant challenges. The detailed modeling of power lines is a prerequisite for detecting potential risks. Thus, this study introduces a fourfold strategy to extract and correctly reconstruct huge gaps in HVTL conductor bundles. In the first step, the power lines are retrieved as span points using the 3-D voxel grid. The second step generates bundle masks by splitting span points into several segments. A bipartite graph connects bundle segments with wide gaps utilizing bundle location information from these masks. Third, bundles are segmented again to form conductor masks for sub-conductor extraction utilizing image-based algorithms and probability. Finally, the mathematical model reconstructs the retrieved power lines as sub-conductors. The proposed technique is tested on 58 spans with different power line configurations from two low-point-density datasets with a few constant parameters. The test results demonstrate the resilience of the proposed method in effectively repairing large gaps in bundles and sub-conductors. The sub-conductors are retrieved with 90% accuracy. The suggested method robustly reconstructs sub-conductors with a fitting residual error of less than 0.07 m.

Utilities have recently shown a considerable interest in extracting powerlines from laser scanning data for periodic utility monitoring. However, if the powerline runs through a forest, extraction is more difficult. A robust and exact power line model is crucial for adequate clearance and detecting potential hazards. Thus, this study establishes and automated method for reconstructing high-voltage bundle conductors. The powerline bundles at various heights are first recovered and divided into segments. The fitted residuals of each segment are used to estimate the number of sub-conductors. The point distance formula and 3D line fitting are used to extract individual sub-conductors. Each part is separated according to its positive and negative distance values. Finally, the sub-conductor segments are reassembled using a random sample consensus (RANSAC) technique. On five datasets, the proposed technique extracts and reconstructs individual sub-conductors with great precision.

Accurate registration of multispectral satellite images is a challenging task due to the significant and nonlinear radiometric differences between these data. To address this problem, this letter explores the strategy of geometrical similarity between triplets of feature points, and it is combined with the structural similarity between images in a feature-based image registration framework. The underlying principle is that the structural and geometrical similarities generally preserve across the images being registered. In this feature-based image registration framework, a set of control points (CPs) are first detected. Then, the geometric similarity between triplets of CPs is defined, followed by a ranking operation of these triplets of CPs. The highly ranked triplets are used to estimate a spatial transformation between images. Finally, initial matches obtained by a benchmark registration technique are refined by the estimated transformation. The experimental results demonstrate the great effectiveness of the proposed technique for registering multispectral satellite images.

This paper introduces an effective method for outlier detection from the point cloud data. Although, the state-of-the-art methods offer good results in removing outliers, in most of the cases inliers are also removed erroneously. This paper focuses on this issue using the information based on a relative location from a point to its neighbours and a robust z-score based on a statistical approach. Synthetic datasets for 3D building roofs have been created to evaluate the performance. When compared with the existing methods, the proposed method exhibits better performance, i.e., 19% more recall for inliers, 6% more precision for outliers and 10% more overall accuracy. In other words, it not only preserves the inliers, but also correctly removes the outliers with a better precision rate than the current state-of-the-arts methods.

Approximation of the geometric features is an essential step in point cloud segmentation and surface reconstruction. Often, the planar surfaces are estimated using principal component analysis (PCA), which is sensitive to noise and smooths the sharp features. Hence, the segmentation results into unreliable reconstructed surfaces. This article presents a point cloud segmentation method for building detection and roof plane extraction. It uses PCA for saliency feature estimation including surface curvature and point normal. However, the point normals around the anisotropic surfaces are approximated using a consistent isotropic sub-neighbourhood by Low-Rank Subspace with prior Knowledge (LRSCPK). The developed segmentation technique is tested using two real-world samples and two benchmark datasets. Per-object and per-area completeness and correctness results indicate the robustness of the approach and the quality of the reconstructed surfaces and extracted buildings.

Automatic 3D extraction of building roofs from remotely sensed data is important for many applications including city modeling. This paper proposes a new method for automatic 3D roof extraction through an effective integration of LIDAR (Light Detection And Ranging) data and multispectral orthoimagery. Using the ground height from a DEM (Digital Elevation Model), the raw LIDAR points are separated into two groups. The first group contains the ground points that are exploited to constitute a ‘ground mask’. The second group contains the non-ground points which are segmented using an innovative image line guided segmentation technique to extract the roof planes. The image lines extracted from the grey-scale version of the orthoimage are classified into several classes such as ‘ground’, ‘tree’, ‘roof edge’ and ‘roof ridge’ using the ground mask and colour and texture information from the orthoimagery. During roof plane extraction the lines from the later two classes are used to fit roof planes to the neighbouring non-ground LIDAR points. Finally, a new rule-based procedure is applied to remove planes constructed on trees. Experimental results show that the proposed method successfully removes vegetation and offers high extraction rates.

Separation of buildings from trees is a major challenge in automatic building detection. In residential and hilly areas, buildings are often surrounded by dense vegetation. This paper presents a three-step method for effective separation of buildings from trees. Firstly, height and width thresholds are applied to LIDAR data for removing small bushes and trees with small horizontal coverage, respectively. The generation of the building mask, where each black region indicates a void area from which there are no laser returns below the height threshold, also helps in separation of buildings from the nearby trees. Then image entropy and colour information are applied together to remove trees exhibiting high texture. Finally, an innovative rule-based procedure is employed using the edge orientation histogram from the imagery to eliminate the remaining trees. Experimental results show that the algorithm offers high building detection rate in complex scenes which are hilly and densely vegetated.

Although a large number of building roof reconstruction techniques from the remotely sensed data have been proposed, there is a serious lack in standard and automatic evaluation system. The manual evaluation systems are based on human judgment and so may be biased. Moreover, the application of such an evaluation system on a large data set requires a considerable amount of human involvement. In addition, the threshold-based evaluation systems that employ one or more thresholds while making correspondences between the detection and reference sets are also subjective since there is no unique way to choose the thresholds. This paper proposes an automatic and threshold-free evaluation system using the centre distances between the detected and reference entities. It neither requires any human involvement nor uses any thresholds. As a result, the proposed evaluation system can be exploited for a bias free evaluation of any roof reconstruction techniques, even on large data sets.

Corner detectors have many applications in computer vision and image identification and retrieval. Contour-based corner detectors directly or indirectly estimate a significance measure (e.g., curvature) on the points of a planar curve, and select the curvature extrema points as corners. While an extensive number of contour-based corner detectors have been proposed over the last four decades, there is no comparative study of recently proposed detectors. This paper is an attempt to fill this gap. The general framework of contour-based corner detection is presented, and two major issues-curve smoothing and curvature estimation, which have major impacts on the corner detection performance, are discussed. A number of promising detectors are compared using both automatic and manual evaluation systems on two large datasets. It is observed that while the detectors using indirect curvature estimation techniques are more robust, the detectors using direct curvature estimation techniques are faster.

The SIFT (scale invariant feature transform) has demonstrated its superior performance in identifying transformed images over many other approaches. However, both of its detection and matching stages are expensive, because a large number of keypoints are detected in the scale-space and each keypoint is described using a 128-dimensional vector. We present two possible solutions for feature-point reduction. First is to down scale the image before the SIFT keypoint detection and second is to use corners (instead of SIFT keypoints) which are visually significant, more robust, and much smaller in number than the SIFT keypoints. Either the curvature descriptor or the highly distinctive SIFT descriptors at corner locations can be used to represent corners.We then describe a new feature-point matching technique, which can be used for matching both the down-scaled SIFT keypoints and corners. Experimental results show that two feature-point reduction solutions combined with the SIFT descriptors and the proposed feature-point matching technique not only improve the computational efficiency and decrease the storage requirement, but also improve the transformed image identification accuracy (robustness).

Geometric transformations change the image pixel positions where the watermark is embedded. As a result, a copyright verifier fails to detect the watermark even though the watermark is present in the transformed image. The content-based watermarking schemes try to solve this problem by first dividing the image into many disjoint patches and then locating those patches with respect to the salient points of the image. Most of the existing content-based schemes are vulnerable to scaling and general affine transformations. They do not work well against high JPEG compression either. This paper presents a content-based blind watermarking scheme using corners. The proposed scheme achieves robustness against affine transformations and signal processing attacks by using triangular patch normalization and spread spectrum watermarking technique respectively.

There are many applications, such as image copyright protection, where transformed images of a given test image need to be identified. The solution to this identification problem consists of two main stages. In stage one, certain representative features, such as corners, are detected in all images. In stage two, the representative features of the test image and the stored images are compared to identify the transformed images for the test image.Curvaturescale-space(CSS) corner detectors look for curvature maxima or inflection points on planar curves. However, the arc-length used to parameterize the planar curves by the existing CSS detectors is not invariant to geometric transformations such as scaling. As a solution to stage one, this paper presents an improved CSS corner detector using the affine-length parameterization which is relatively invariant to affine transformations. We then present an improved corner matching technique as a solution to the stage two. Finally, we apply the proposed corner detection and matching techniques to identify the transformed images for a given image and report the promising results.

Many contour-based image corner detectors are based on the curvature scale-space (CSS). We identify the weaknesses of the CSS-based detectors. First, the ldquocurvaturerdquo itself by its ldquodefinitionrdquo is very much sensitive to the local variation and noise on the curve, unless an appropriate smoothing is carried out beforehand. In addition, the calculation of curvature involves derivatives of up to second order, which may cause instability and errors in the result. Second, the Gaussian smoothing causes changes to the curve and it is difficult to select an appropriate smoothing-scale, resulting in poor performance of the CSS corner detection technique. We propose a complete corner detection technique based on the chord-to-point distance accumulation (CPDA) for the discrete curvature estimation. The CPDA discrete curvature estimation technique is less sensitive to the local variation and noise on the curve. Moreover, it does not have the undesirable effect of the Gaussian smoothing. We provide a comprehensive performance study. Our experiments showed that the proposed technique performs better than the existing CSS-based and other related methods in terms of both average repeatability and localization error.

Curvature scale-space (CSS) corner detectors look for curvature maxima or inflection points on planar curves. They use arc-length parameterized curvature. Therefore, they are not robust to affine transformations since the arc-length of a curve is not preserved under affine transformations. However, the affine-length of a curve is relatively invariant to affine transformations. This paper presents an improved CSS corner detector by applying the affine-length parameterized curvature to the CSS corner detection technique. A thorough robustness study has been carried out on a large database considering a wide range of affine transformations.

In contour-based comer detectors, absolute curvature values of many corners either remain unaltered or change slightly under affine transformations. Moreover, affine-length of a curve is relatively invariant to affine transformations. This paper presents a novel corner matching technique using corners detected by contour-based detectors. For each corner we use its position, absolute curvature value, and affine-lengths between this corner and other corners on the same curve. The iterative matching procedure tries to find three corner matches with minimum absolute curvature difference to calculate affine transformation parameters. Original corners are transformed with the estimated parameters prior to matching with the test corner set.

The most significant bit (MSB)-plane of an image is least likely to change by the most signal processing operations. Watermarking techniques are, however, unable to exploit the MSB-plane, as embedding any information there introduces the highest distortion. This paper presents a novel rotation, scale, and translation (RST)-resistant multi-bit logo-based copyright protection scheme using the most significant gray-scale bits at the region-of-interest, automatically selected by the invariant centroid (1C) of the image. How the RST-attack can be reversed using the 1C and geometric moments has been proposed. During verification, the test image is restored to its approximate original through reversing any possible RST attack before calculating signature. To avoid any bias, a new MSB-based attack has also been proposed. Experimental results have clearly demonstrated the superiority of the proposed scheme.

The performance of existing copyright protection schemes is questionable due to their vulnerability to geometric transformations. Though a few of them can resist global geometric transformations like rotation and scaling attacks, most of them are vulnerable to rotation-scale and cropping attacks. This paper presents a novel geometric distortion correction scheme robust to global geometric transformations. It restores an attacked image to its approximate original by reversing the attack using the invariant centroid and geometric moments of the image. Experimental results show the effectiveness of the proposed scheme