From instance segmentation to physical quantification: High-resolution UAV-based dataset for façade defect assessment

With the development of satellite technology, up to date imaging mode of synthetic aperture radar (SAR) satellite can provide higher resolution SAR imageries, which benefits ship detection and instance segmentation. Meanwhile, object detectors based on convolutional neural network (CNN) show high performance on SAR ship detection even without land-ocean segmentation; but with respective shortcomings, such as the relatively small size of SAR images for ship detection, limited SAR training samples, and inappropriate annotations, in existing SAR ship datasets, related research is hampered. To promote the development of CNN based ship detection and instance segmentation, we have constructed a High-Resolution SAR Images Dataset (HRSID). In addition to object detection, instance segmentation can also be implemented on HRSID. As for dataset construction, under the overlapped ratio of 25%, 136 panoramic SAR imageries with ranging resolution from 1m to 5m are cropped to $800 \times 800$ pixels SAR images. To reduce wrong annotation and missing annotation, optical remote sensing imageries are applied to reduce the interferes from harbor constructions. There are 5604 cropped SAR images and 16951 ships in HRSID, and we have divided HRSID into a training set (65% SAR images) and test set (35% SAR images) with the format of Microsoft Common Objects in Context (MS COCO). 8 state-of-the-art detectors are experimented on HRSID to build the baseline; MS COCO evaluation metrics are applicated for comprehensive evaluation. Experimental results reveal that ship detection and instance segmentation can be well implemented on HRSID.

For defense applications, rapid X-ray inspection of propellant samples is essential for the identification and assessment of defects. Automation of this process using artificial intelligence is possible by properly training a neural network model. Convolution Neural Networks (CNNs) have recently demonstrated excellent success in both the tasks of image recognition and localisation using an adequate amount of data. In real-world, it’s not an easy task to produce the correct amount of experimental data required for the deep neural network to operate. In this work, we propose a method for producing synthetic radiographic data that is supported by ray tracing based radiographic simulations for the deep learning algorithms to automatically detect anomaly in X-ray images. The simulation results, which are then supplemented by noise extracted from the experimental data, show a good comparison with the measurements. This Simulation assisted Automatic Defect Recognition (Sim-ADR) system simultaneously perform defect detection and defect instance segmentation. The accuracy of the defect detection system is more than 87% on a testing set included 416 images.

Adaptive wind turbine blade surface defect detection based on deep learning and transfer learning

Prototyping and evaluation of a novel machine vision system for real-time, automated quality grading of sweetpotatoes

With the development of synthetic aperture radar (SAR) system technology and the wide application of deep learning technology, ship detection on SAR images has rapidly developed. Benefit from the strong generalization ability and end-to-end training capabilities, convolution neural network (CNN) based ship detection methods have the proprietary advantage in high-performance SAR ship detection. However, relevant SAR ship detection methods adopt rectangular bounding boxes to locate the ships which are unable to extract the contours feature of the ships. To solve this problem, we proposed a precise instance segmentation network for high-resolution SAR images. The method combines the bottom-up path augmentation module, global context module, and soft non-maximum suppression to improve Mask R-CNN for segmenting high-resolution SAR ships in pixel-wise. The network is trained and tested on the high-resolution SAR images dataset (HRSID), and the ablation experiments are conducted with Microsoft Common Objects in Context (MS COCO) evaluation metrics to verify the effects of each module. Quantitatively, the experimental results show that the method exceeds vanilla Mask R-CNN 2% AP in instance segmentation of high-resolution SAR images. Meanwhile, the visualized instance segmentation results indicate that our method fits the practical application, and it possesses the ability to extract the contour of the ship, which is more conducive to the instance segmentation of SAR images.

The instance segmentation of ships in synthetic aperture radar (SAR) images aims to interpret detailed position and shape information, holding significant potential applications in ocean-going ship monitoring and port scheduling. Existing SAR ship instance segmentation methods face challenges such as expensive label production costs, weak edge detail perception, and insufficient adaptation to intrinsic limitations in SAR images, such as object information loss and speckle noise. Addressing these challenges, we propose a gradient prior guided and SAR image adaptation enhanced semi-supervised instance segmentation (GGSE-SSIS) method. This method, rooted in a teacher-student framework, leverages pseudo-labels generated by a teacher model trained on a small amount of data to guide the student model toward mask prediction, thus achieving high-performance instance segmentation of SAR ships at low annotation costs. We have also meticulously designed a gradient prior guidance (GPG) module to enhance the gradient consistency between the objects and the corresponding mask proposals, facilitating the perception of target edge details. Additionally, the SAR image adaptation enhancement (SIAE) operation is introduced into the GGSE-SSIS method to construct more robust training signals while enhancing adaptability to intrinsic limitations such as object information loss and speckle noise in SAR images. Experimental results on high-resolution SAR images dataset (HRSID) and polygon segmentation SAR ship detection dataset (PSeg-SSDD) demonstrate that the proposed GGSE-SSIS achieves segmentation performance close to that of fully supervised methods using only 30% pixel-level annotations, effectively balancing annotation costs and visual perceptual effects in the SAR ship instance segmentation task.

Deep convolutional neural networks (DCNN)-based methods have been applied widely to ship detection in SAR images. However, most DCNN-based ship target detectors that focus on the detection performance ignore the computation complexity. We propose a lightweight anchor-free ship detection network (LASDNet) for SAR images to tackle this problem. First, a lightweight backbone utilizing a double fusion with squeeze-and-excitation-bottleneck block under the CSPNet design (CSP-DFSEB) and three pooling blocks (i.e., EVE, FCT, and ME blocks) are constructed, which achieves a balance between accuracy and efficiency. Second, a transformer-based aggregation layer conducts feature fusion. Finally, an improved one-stage anchor-free detector FCOS is presented. The analyses of the High-Resolution SAR Images Dataset for Ship Detection and Instance Segmentation (HRSID) dataset show that the proposed detector has the second least number of parameters (1.15 MB), the lowest computation complexity (1.01 GFLOPs), and the highest average precision (59.25) compared with other state-of-the-art methods.

Deep learning (DL) based synthetic aperture radar (SAR) imagery ship detection is challenged by multiscale ships on the identical SAR image, which inevitably leads to insufficient and low-quality positive samples during training and ultimately degrades detection performance. To address this issue, we propose a Scale-Sensitive Adaptive Sample Allocation Strategy (SSA-SAS) for SAR ship detection. SSA-SAS ranks candidate boxes using a unified score that integrates a scale-sensitive Wasserstein distance (SSWD), a shape cost, and classification confidence. SSWD serves as the core regression metric, enabling adaptive tolerance to positional offsets based on object scale. Meanwhile, the shape cost introduces morphological priors to guide early-stage optimization. These components jointly enhance the quantity and quality of selected positive samples throughout training. Experimental results show that SSA-SAS improves average precision (AP) by up to 2.6% on the high-resolution SAR images dataset for ship detection and instance segmentation (HRSID) dataset and 1.4% on the SAR ship detection dataset (SSDD), while accelerating network convergence by approximately 5.0%.

Deep learning models can enhance the detection efficiency and accuracy of rapid on-site screening for imported grains at customs, satisfying the need for high-throughput, efficient, and intelligent operations. However, the construction of datasets, which is crucial for deep learning models, often involves significant labor and time costs. Addressing the challenges associated with establishing high-resolution instance segmentation datasets for small objects, we integrate two zero-shot models, Grounding DINO and Segment Anything model, into a dataset annotation pipeline. Furthermore, we encapsulate this pipeline into a software tool for manual calibration of mislabeled, missing, and duplicated annotations made by the models. Additionally, we propose preprocessing and postprocessing methods to improve the detection accuracy of the model and reduce the cost of subsequent manual correction. This solution is not only applicable to rapid screening for quarantine weeds, seeds, and insects at customs but can also be extended to other fields where instance segmentation is required.

The growth of the most significant field crops such as rice, wheat, maize, and soybean are influenced because of various pests. And crop production is decreased due to various categories of insects. Deep learning technologies significantly increased the efficiency of identifying and controlling agricultural pests attack. However, agricultural pests images obtained are often obscure and unclear because of the sparse density of cameras deployed in the real farmland. This always makes pests difficult to recognize and monitor. Additionally, the existing classification and segmentation methods are not satisfying for the identification of low-resolution images because they are pre-trained on the clear and high-resolution datasets. Therefore, it is crucial to restore and upscale the obtained low-resolution pest images in order to improve classification accuracy and the recall rate of the instance segmentation. In this paper, we propose a generative adversarial network (GAN) with quadra-attention and residual and dense fusion mechanisms to transform low-resolution pest images. Compared with previous state-of-the-art PSNR-oriented super-resolution methods, our proposed method is more powerful in image reconstruction and achieves the state of the art performance. The experiment results show that after reconstructing with our proposed gan, the recall rate increased by 182.89% and classification accuracy also improved a lot. Besides, our proposed method could decrease the density of the camera layout in the agricultural Internet of Things (IOT) monitor systems and the cost of infrastructure, which is practical for real-world applications.

A fast and accurate deep learning method for strawberry instance segmentation

PollenBB16 is an RGB pollen image dataset of Chilean flora with pixel-accurate instance segmentation masks, whose annotation was fully verified by an expert palynologist to guarantee the taxonomic reliability of every published instance. The dataset is designed to close a concrete gap in existing palynological datasets, which typically combine low taxonomic diversity, few samples per class, and low-resolution crops restricted to bounding boxes. PollenBB16 contains 16,198 brightfield optical microscopy images at the native resolution of 3088 × 2064 pixels and 36,383 pixel-accurate polygons across 16 species from the Biobío Region, spanning endemic, native and exotic species of high ecological and melliferous value such as Eucryphia glutinosa and Quillaja saponaria (endemic), Gevuina avellana and Aristotelia chilensis (native), and Medicago sativa and Brassica rapa (introduced). Each spatial position is recorded at three focal planes. The displacement along the z axis reveals features of the exine together with information on the internal structure of the grain that remain inaccessible on a single plane. From this multifocal information, more robust convolutional networks can be trained with more accurate classification. The operational quality of the dataset is backed by a leakage-safe partition that keeps the three focal planes of the same position in the same subset to avoid metric inflation, complemented by a YOLO11n-seg baseline trained for 50 epochs that reaches 0.985 mask mAP@50 on the validation set, establishing a reproducible reference point. Beyond deep learning, PollenBB16 enables interdisciplinary applications in aerobiology, biodiversity monitoring under climate change, ecological restoration of the South American temperate forest, and botanical-origin authentication of Chilean monofloral honeys.

This article proposes a mask refinement method for chromosome instance segmentation. The proposed method exploits the knowledge representation capability of Neural Knowledge DNA (NK-DNA) to capture the semantics of the chromosome’s shape, texture, and key points, and then it uses the captured knowledge to improve the accuracy and smoothness of the masks. We validate the method’s effectiveness on our latest high-resolution chromosome image dataset. The experimental results show that our proposed method’s mask average precision (MaskAP) is 3.66% higher than Mask R-CNN and outperforms advanced Cascade Mask R-CNN by 1.35%.

Existing deep learning (DL)-based synthetic aperture radar (SAR) ship instance segmentation models mostly extract feature subsets at the single level of feature pyramid network (FPN), and also ignore context information of the region of interest (ROI), which both hinder accuracy improvements. Thus, a full-level context squeeze-and-excitation ROI extractor (FL-CSE-ROIE) is proposed to handle these problems. FL-CSE-ROIE has three novelties: 1) full-level, i.e., extract feature subsets at each level of FPN to retain multi-scale features; 2) context, i.e., add multi-context surroundings of different scopes to ROIs to ease background interferences; and 3) squeeze-and-excitation (SE), i.e., balance contributions of different scope contexts to highlight valuable features and suppress useless ones. FL-CSE-ROIE is applied to the fashionable hybrid task cascade (HTC) model. Results on two open SAR ship detection dataset (SSDD) and high-resolution SAR images dataset (HRSID) confirm its effectiveness. Moreover, another two improvements to HTC are also proposed to enhance accuracy further: 1) the raw deconv is replaced with a content-aware reassembly of features block (CARAFEB) to enable larger receptive fields and 2) the raw <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$1\times1$ </tex-math></inline-formula> conv for the mask information interaction is replaced with a global feature self-attention block (GFSAB) to enhance interaction benefits. Finally, FL-CSE-ROIE surpasses the other nine advanced models, better than the suboptimal model by 2.4%/2.3% detection average precision (AP) and 3.0%/2.5% segmentation AP on SSDD/HRSID.

This dataset is an expanded version of a previously published collection of high-resolution RGB images of Urochloa spp. genotypes, initially designed to facilitate automated classification of phenological stages and raceme identification in forage breeding trials. The original dataset included 2400 images of 200 genotypes captured under controlled conditions, supporting the development of computer vision models for High-Throughput Phenotyping (HTP). In this updated release, 139 additional images and 24,983 new annotations have been added, bringing the dataset to a total of 2539 images and 47,323 raceme annotations. This version introduces increased diversity in image-capture conditions, with data collected from two geographic locations (Palmira, Colombia, and Ocozocoautla de Espinosa, Mexico) and a range of image-capture devices, including smartphones (e.g. Realme C53 and Oppo Reno 11), a Nikon D5600 camera, and a Phantom 4 Pro V2 drone. Images now vary in perspective (nadir, high-angle, and frontal) and capture distance (1–3 meters), enhancing the dataset applicability for robust Deep Learning (DL) models. Compared to the original dataset, raceme density per plant has nearly doubled in some samples, offering higher raceme overlap for advanced instance segmentation tasks. This expanded dataset supports deeper exploration of phenotypic variation in Urochloa spp. and offers greater potential for developing adaptable models in crop phenotyping.