Large Set Of Images Research Articles

CVINet: A deep learning based model for the diagnosis of chronic venous insufficiency in lower extremity using infrared thermal images

AbstractChronic venous insufficiency (CVI) is a venous disorder characterized by impaired blood flow from the lower extremities back to the heart, leading to various symptoms and complications. Accurate and timely diagnosis of CVI is critical for effective management and prevention of further complications. Deep learning (DL) models have shown results that are promising in the field of medical image analysis tasks over the past few years. This article presents CVINet, a enhanced DL‐based convolutional neural network (CNN) model that has been developed specifically for the purpose of diagnosing CVI utilizing thermal imaging data. The proposed CNN model extracts relevant features from input images using a multi‐layered architecture with convolutional layers and pooling layers. Following feature extraction, data are directed to fully connected layers for performing classification task. To enhance the performance of this model, we employ techniques such as dropout regularization and batch normalization. This model utilizes its ability to automatically learn discriminative features from raw thermal images. A large set of thermal images from both CVI patients and healthy individuals was acquired for training and testing of the model. A comparison between few conventional pre‐trained models was conducted, and the best performing model for classification was examined. Pre‐trained VGG‐16, EfficientNet‐B0, and ResNet‐152 models produced classification accuracies of 94.7%, 95.3%, and 95.8%, respectively. The Proposed CVINet model attained a classification accuracy of 96.8% on binary classification, demonstrating high performance compared with other state‐of‐the‐art networks. Overall, our proposed CNN model offers a promising solution for accurately classifying CVI and normal subjects based on thermal images which outperforms the diagnosis by clinician. The potential benefits include early detection of CVI and personalized treatment strategies that can improve patient outcomes. They improve the accuracy of patient diagnosis, reducing radiologist workload and providing them with a tool that can automatically assess the condition of the lower limb venous condition, lowering the risk of misdiagnosis. Furthermore, improving healthcare service quality and early detection can alter the course of the disease and reduce the morbidity and mortality rate of the disease.

Image factory: A method for synthesizing novel CT images with anatomical guidance.

Deep learning in medical applications is limited due to the low availability of large labeled, annotated, or segmented training datasets. With the insufficient data available for model training comes the inability of these networks to learn the fine nuances of the space of possible images in a given medical domain, leading to the possible suppression of important diagnostic features hence making these deep learning systems suboptimal in their performance and vulnerable to adversarialattacks. We formulate a framework to address this lack of labeled data problem. We test this formulation in computed tomographic images domain and present an approach that can synthesize large sets of novel CT images at high resolution across the full Hounsfield (HU)range. Our method only requires a small annotated dataset of lung CT from 30 patients (available online at the TCIA) and a large nonannotated dataset with high resolution CT images from 14k patients (received from NIH, not publicly available). It then converts the small annotated dataset into a large annotated dataset, using a sequence of steps including texture learning via StyleGAN, label learning via U-Net and semi-supervised learning via CycleGAN/Pixel-to-Pixel (P2P) architectures. The large annotated dataset so generated can then be used for the training of deep learning networks for medical applications. It can also be put to use for the synthesis of CT images with varied anatomies that were nonexistent within either of the input datasets, enriching the dataset evenfurther. We demonstrate our framework via lung CT-Scan synthesis along with their novel generated annotations and compared it with other state of the art generative models that only produce images without annotations. We evaluate our framework effectiveness via a visual turing test with help of a few doctors and radiologists. We gain the capability of generating an unlimited amount of annotated CT images. Our approach works for all HU windows with minimal depreciation in anatomical plausibility and hence could be used as a general purpose framework for annotated data augmentation for deep learning applications in medical imaging.

Semi-Supervised Semantic Segmentation-Based Remote Sensing Identification Method for Winter Wheat Planting Area Extraction

To address the cost issue associated with pixel-level image annotation in fully supervised semantic segmentation, a method based on semi-supervised semantic segmentation is proposed for extracting winter wheat planting areas. This approach utilizes self-training with pseudo-labels to learn from a small set of images with pixel-level annotations and a large set of unlabeled images, thereby achieving the extraction. In the constructed initial dataset, a random sampling strategy is employed to select 1/16, 1/8, 1/4, and 1/2 proportions of labeled data. Furthermore, in conjunction with the concept of consistency regularization, strong data augmentation techniques are applied to the unlabeled images, surpassing classical methods such as cropping and rotation to construct a semi-supervised model. This effectively alleviates overfitting caused by noisy labels. By comparing the prediction results of different proportions of labeled data using SegNet, DeepLabv3+, and U-Net, it is determined that the U-Net network model yields the best extraction performance. Moreover, the evaluation metrics MPA and MIoU demonstrate varying degrees of improvement for semi-supervised semantic segmentation compared to fully supervised semantic segmentation. Notably, the U-Net model trained with 1/16 labeled data outperforms the models trained with 1/8, 1/4, and 1/2 labeled data, achieving MPA and MIoU scores of 81.63%, 73.31%, 82.50%, and 76.01%, respectively. This method provides valuable insights for extracting winter wheat planting areas in scenarios with limited labeled data.

MIGGRI: A multi-instance graph neural network model for inferring gene regulatory networks for Drosophila from spatial expression images.

Recent breakthrough in spatial transcriptomics has brought great opportunities for exploring gene regulatory networks (GRNs) from a brand-new perspective. Especially, the local expression patterns and spatio-temporal regulation mechanisms captured by spatial expression images allow more delicate delineation of the interplay between transcript factors and their target genes. However, the complexity and size of spatial image collections pose significant challenges to GRN inference using image-based methods. Extracting regulatory information from expression images is difficult due to the lack of supervision and the multi-instance nature of the problem, where a gene often corresponds to multiple images captured from different views. While graph models, particularly graph neural networks, have emerged as a promising method for leveraging underlying structure information from known GRNs, incorporating expression images into graphs is not straightforward. To address these challenges, we propose a two-stage approach, MIGGRI, for capturing comprehensive regulatory patterns from image collections for each gene and known interactions. Our approach involves a multi-instance graph neural network (GNN) model for GRN inference, which first extracts gene regulatory features from spatial expression images via contrastive learning, and then feeds them to a multi-instance GNN for semi-supervised learning. We apply our approach to a large set of Drosophila embryonic spatial gene expression images. MIGGRI achieves outstanding performance in the inference of GRNs for early eye development and mesoderm development of Drosophila, and shows robustness in the scenarios of missing image information. Additionally, we perform interpretable analysis on image reconstruction and functional subgraphs that may reveal potential pathways or coordinate regulations. By leveraging the power of graph neural networks and the information contained in spatial expression images, our approach has the potential to advance our understanding of gene regulation in complex biological systems.

Sparse-view synchrotron X-ray tomographic reconstruction with learning-based sinogram synthesis.

Synchrotron radiation can be used as a light source in X-ray microscopy to acquire a high-resolution image of a microscale object for tomography. However, numerous projections must be captured for a high-quality tomographic image to be reconstructed; thus, image acquisition is time consuming. Such dense imaging is not only expensive and time consuming but also results in the target receiving a large dose of radiation. To resolve these problems, sparse acquisition techniques have been proposed; however, the generated images often have many artefacts and are noisy. In this study, a deep-learning-based approach is proposed for the tomographic reconstruction of sparse-view projections that are acquired with a synchrotron light source; this approach proceeds as follows. A convolutional neural network (CNN) is used to first interpolate sparse X-ray projections and then synthesize a sufficiently large set of images to produce a sinogram. After the sinogram is constructed, a second CNN is used for error correction. In experiments, this method successfully produced high-quality tomography images from sparse-view projections for two data sets comprising Drosophila and mouse tomography images. However, the initial results for the smaller mouse data set were poor; therefore, transfer learning was used to apply the Drosophila model to the mouse data set, greatly improving the quality of the reconstructed sinogram. The method could be used to achieve high-quality tomography while reducing the radiation dose to imaging subjects and the imaging time and cost.

African bovid tribe classification using transfer learning and computer vision.

Objective analytical identification methods are still a minority in the praxis of paleobiological sciences. Subjective interpretation of fossils and their modifications remains a nonreplicable expert endeavor. Identification of African bovids is a crucial element in the reconstruction of paleo-landscapes, ungulate paleoecology, and, eventually, hominin adaptation and ecosystemic reconstruction. Recent analytical efforts drawing on Fourier functional analysis and discrimination methods applied to occlusal surfaces of teeth have provided a highly accurate framework to correctly classify African bovid tribes and taxa. Artificial intelligence tools, like computer vision, have also shown their potential to be objectively more accurate in the identification of taphonomic agency than human experts. For this reason, here we implement some of the most successful computer vision methods, using transfer learning and ensemble analysis, to classify bidimensional images of African bovid teeth and show that 92% of the large testing set of images of African bovid tribes analyzed could be correctly classified. This brings an objective tool to paleoecological interpretation, where bovid identification and paleoecological interpretation can be more confidently carried out.

Deep Learning for Fully Automatic Tumor Segmentation on Serially Acquired Dynamic Contrast-Enhanced MRI Images of Triple-Negative Breast Cancer.

Accurate tumor segmentation is required for quantitative image analyses, which are increasingly used for evaluation of tumors. We developed a fully automated and high-performance segmentation model of triple-negative breast cancer using a self-configurable deep learning framework and a large set of dynamic contrast-enhanced MRI images acquired serially over the patients' treatment course. Among all models, the top-performing one that was trained with the images across different time points of a treatment course yielded a Dice similarity coefficient of 93% and a sensitivity of 96% on baseline images. The top-performing model also produced accurate tumor size measurements, which is valuable for practical clinical applications.

Harnessing CRNN Technology for Image-Based Sequence Recognition

We are studying the challenge of finding text in everyday images using advanced computer techniques, specifically deep learning. Our system can accurately spot and read text in images. It uses Convolutional Recurrent Neural Networks (CRNNs), a type of deep learning model, to locate and understand the text. Instead of relying on manual rules, our model learns from the data to identify text patterns. Our projects unique feature is that our model is trained on single- word images, making it versatile in recognizing different text styles and backgrounds. We trained our model using a large dataset, including the “MJ Synthetic Word Dataset”, which consists of 50,000 out of which we have taken 35000 astraining images, 15000 as validation images. The results have been promising, indicating potential real-world applications such as searching for words in large sets of images. Key Words: Advanced computer techniques, Deep learning, System, Spot and read text, Convolutional Neural Networks (CRNNs), Deep learning model

Amplifying the Effects of Contrast Agents on Magnetic Resonance Images Using a Deep Learning Method Trained on Synthetic Data.

Artificial intelligence (AI) methods can be applied to enhance contrast in diagnostic images beyond that attainable with the standard doses of contrast agents (CAs) normally used in the clinic, thus potentially increasing diagnostic power and sensitivity. Deep learning-based AI relies on training data sets, which should be sufficiently large and diverse to effectively adjust network parameters, avoid biases, and enable generalization of the outcome. However, large sets of diagnostic images acquired at doses of CA outside the standard-of-care are not commonly available. Here, we propose a method to generate synthetic data sets to train an "AI agent" designed to amplify the effects of CAs in magnetic resonance (MR) images. The method was fine-tuned and validated in a preclinical study in a murine model of brain glioma, and extended to a large, retrospective clinical human data set. A physical model was applied to simulate different levels of MR contrast from a gadolinium-based CA. The simulated data were used to train a neural network that predicts image contrast at higher doses. A preclinical MR study at multiple CA doses in a rat model of glioma was performed to tune model parameters and to assess fidelity of the virtual contrast images against ground-truth MR and histological data. Two different scanners (3 T and 7 T, respectively) were used to assess the effects of field strength. The approach was then applied to a retrospective clinical study comprising 1990 examinations in patients affected by a variety of brain diseases, including glioma, multiple sclerosis, and metastatic cancer. Images were evaluated in terms of contrast-to-noise ratio and lesion-to-brain ratio, and qualitative scores. In the preclinical study, virtual double-dose images showed high degrees of similarity to experimental double-dose images for both peak signal-to-noise ratio and structural similarity index (29.49 dB and 0.914 dB at 7 T, respectively, and 31.32 dB and 0.942 dB at 3 T) and significant improvement over standard contrast dose (ie, 0.1 mmol Gd/kg) images at both field strengths. In the clinical study, contrast-to-noise ratio and lesion-to-brain ratio increased by an average 155% and 34% in virtual contrast images compared with standard-dose images. Blind scoring of AI-enhanced images by 2 neuroradiologists showed significantly better sensitivity to small brain lesions compared with standard-dose images (4.46/5 vs 3.51/5). Synthetic data generated by a physical model of contrast enhancement provided effective training for a deep learning model for contrast amplification. Contrast above that attainable at standard doses of gadolinium-based CA can be generated through this approach, with significant advantages in the detection of small low-enhancing brain lesions.

Marine Environmental Impact on CFAR Ship Detection as Measured by Wave Age in SAR Images

Satellite synthetic aperture radar (SAR) images are recognized as one of the most efficient tools for day/night, all weather and large area monitoring of ships at sea. However, false alarms discrimination is still one key problem on SAR ship detection. While many discrimination techniques have been proposed for the treatment of false alarms, not enough emphasis has been targeted to explore how obtained false alarms are related to the changing ocean environmental conditions. To this end, we combined a large set of Sentinel-1 SAR images with ocean surface wind and wave data into one dataset. SAR images were separated into three distinct groups according to wave age (WA) conditions present during image acquisition: young wind sea, old wind sea, and swell. A constant false alarm rate (CFAR) ship detection algorithm was implemented based on the generalized gamma distribution (GΓD). Kolmogorov–Smirnov distance was used to analyze the distribution goodness-of-fit among distinct ocean environments. A backscattering analysis of different sizes of ship targets and sea clutter was further performed using the OpenSARShip and automatic identification system (AIS) datasets to assess its separability. We derived a discrimination threshold adjustment based on WA conditions and showed its efficacy to drastically reduce false alarms. To our present knowledge, the use of WA as part of the CFAR and for the adjustment of the threshold of detection is a novelty which could be tested and evaluated for different SAR sensors.

Improved Segmentation of Pulmonary Nodules Using Soft Computing Techniques with SegNet and Adversarial Networks

Lung cancer is seen as one of the most common lung diseases. For the patients having symptoms, the presence of lung nodules is checked by using various imaging techniques. Pulmonary nodules are detected in most of the cases having symptoms. But identifying the type of the nodule and the categorization still remains as a challenge. After confirming the presence of a nodule (benign or malignant) it takes several other steps to identify its characteristics. Improved imaging methods produce results within a short span of time. Research works are being conducted to increase the overall efficiency of the system. The proposed system considers authentic data sources for the study. The benign and malignant samples are considered for the generation of realistic large image sets. The generation of a large data set with the help of a generative adversarial network (GAN) is the first part of the work. The generated images using GAN cannot be differentiated from the original images even by a trained radiologist. This proves the importance of images generated using GAN. A GAN is able to generate 1024 × 1024 resolutions for natural images. Real data images are used to finetune the SegNet output. Through transfer learning, these weights are transferred to the system for segmentation of the images. The training process use real and generated images, which improve theefficiency of the network. The original data from LUNA 16 was used to further generate benign and malignant samples using GAN. A total of 440 images and their augmented images were used for training the GAN, and it generated 1,001,000 images. Hence the overall efficiency of the system was improved. To verify the results, the same various combinations and methods were considered and tabulated with various parameters. Methods with SegNet, GAN, and other combinations were evaluated to verify the efficiency of the system. Receiver operating characteristics were also plotted and compared with the area under the curve for verification of the results.

A hybrid CNN-LSTM model for diagnosing rice nutrient levels at the rice panicle initiation stage

Nitrogen (N) and potassium (K) are two key mineral nutrient elements involved in rice growth. Accurate diagnosis of N and K status is very important for the rational application of fertilizers at a specific rice growth stage. Therefore, we propose a hybrid model for diagnosing rice nutrient levels at the early panicle initiation stage (EPIS), which combines a convolutional neural network (CNN) with an attention mechanism and a long short-term memory network (LSTM). The model was validated on a large set of sequential images collected by an unmanned aerial vehicle (UAV) from rice canopies at different growth stages during a two-year experiment. Compared with VGG16, AlexNet, GoogleNet, DenseNet, and inceptionV3, ResNet101 combined with LSTM obtained the highest average accuracy of 83.81% on the dataset of Huanghuazhan (HHZ, an indica cultivar). When tested on the datasets of HHZ and Xiushui 134 (XS134, a japonica rice variety) in 2021, the ResNet101-LSTM model enhanced with the squeeze-and-excitation (SE) block achieved the highest accuracies of 85.38 and 88.38%, respectively. Through the cross-dataset method, the average accuracies on the HHZ and XS134 datasets tested in 2022 were 81.25 and 82.50%, respectively, showing a good generalization. Our proposed model works with the dynamic information of different rice growth stages and can efficiently diagnose different rice nutrient status levels at EPIS, which are helpful for making practical decisions regarding rational fertilization treatments at the panicle initiation stage.

Shape-selective processing in deep networks: integrating the evidence on perceptual integration

Understanding how deep neural networks resemble or differ from human vision becomes increasingly important with their widespread use in Computer Vision and as models in Neuroscience. A key aspect of human vision is shape: we decompose the visual world into distinct objects, use cues to infer their 3D geometries, and can group several object parts into a coherent whole. Do deep networks use the shape of objects similarly when they classify images? Research on this question has yielded conflicting results, with some studies showing evidence for shape selectivity in deep networks, while others demonstrated clear deficiencies. We argue that these conflicts arise from differences in experimental methods: whether studies use custom images in which only some features are available, images in which different features compete, image pairs that vary along different feature dimensions, or large sets of images to assess how representations vary overall. Each method offers a different, partial view of shape processing. After comparing their advantages and pitfalls, we propose two hypotheses that can reconcile previous results. Firstly, deep networks are sensitive to local, but not global shape. Secondly, the higher layers of deep networks discard some of the shape information that the lower layers are sensitive to. We test these hypotheses by comparing network representations for natural images and silhouettes in which local or global shape is degraded. The results support both hypotheses, but for different networks. Purely feed-forward convolutional networks are unable to integrate shape globally. In contrast, networks with residual or recurrent connections show a weak selectivity for global shape. This motivates further research into recurrent architectures for perceptual integration.

Image Hashing Robust Against Cropping and Rotation

Image recognition is an important mechanism used in various scenarios. In the context of multimedia forensics, its most significant task is to automatically detect already known child and adolescent pornography in a large set of images. When fighting disinformation, it is used to identify images taken out of context or image montages. For this purpose, numerous methods based on robust hashing and feature extraction are already known, and recently also supported by machine learning. However, in general, these methods are either only partially robust to changes such as rotation and pruning, or they require a large amount of data and computation. We present a method based on a simple block hash that is efficient to compute and memory efficient. To be robust against cropping and rotation, we combine the method with image segmentation and a method to normalize the rotation of the objects. Our evaluation shows that the method produces results comparable to much more complex approaches, but requires fewer resources.

Low-level tuning biases in higher visual cortex reflect the semantic informativeness of visual features.

Representations of visual and semantic information can overlap in human visual cortex, with the same neural populations exhibiting sensitivity to low-level features (orientation, spatial frequency, retinotopic position) and high-level semantic categories (faces, scenes). It has been hypothesized that this relationship between low-level visual and high-level category neural selectivity reflects natural scene statistics, such that neurons in a given category-selective region are tuned for low-level features or spatial positions that are diagnostic of the region's preferred category. To address the generality of this "natural scene statistics" hypothesis, as well as how well it can account for responses to complex naturalistic images across visual cortex, we performed two complementary analyses. First, across a large set of rich natural scene images, we demonstrated reliable associations between low-level (Gabor) features and high-level semantic categories (faces, buildings, animate/inanimate objects, small/large objects, indoor/outdoor scenes), with these relationships varying spatially across the visual field. Second, we used a large-scale functional MRI dataset (the Natural Scenes Dataset) and a voxelwise forward encoding model to estimate the feature and spatial selectivity of neural populations throughout visual cortex. We found that voxels in category-selective visual regions exhibit systematic biases in their feature and spatial selectivity, which are consistent with their hypothesized roles in category processing. We further showed that these low-level tuning biases are not driven by selectivity for categories themselves. Together, our results are consistent with a framework in which low-level feature selectivity contributes to the computation of high-level semantic category information in the brain.

Potential landslides risk assessment based on remote sensing technique-a cause study from Dorset, southern England, United Kingdom

This paper presents a method of potential landslide risk assessment from the Dorset Southern England, United Kingdom. It concentrated on potential geological hazards and vulnerability to build a quantitative landslide risk framework in Dorset. In this study, the landslides risk assessment is centered on three components: identification of risks, calculation of risks, and evaluation of risks. Therefore, different natural environmental situations were collected using a large set of optical high-resolution satellite images and remote sensing to determine the probability of landslide occurrence and provide the quantitative model by combining the remote sensing technique. The result shows Dorset area risk level is high. The possibility of landslides mainly comes from natural conditions and human actions. The calculation of the possibility of potential landslides is based on the Bayes’ theorem. Moreover, the limitation of the probability potential risk calculation is insensitivity. The challenge is related to the temporal resolution of precipitation information.

Object detection in hospital facilities: A comprehensive dataset and performance evaluation

Detecting objects in hospital indoor environments is critical for scene understanding and can have various applications in healthcare. Deep learning algorithms have proven to be effective in object recognition from images or videos, but the availability of annotated datasets plays a crucial role in their successful application. However, there is a shortage of datasets for object detection in hospital settings, hindering the advancement of hospital indoor object detection algorithms. In this paper, we present the Hospital Indoor Object Detection (HIOD) dataset, consisting of 4,417 images covering 56 object categories. The HIOD dataset represents the frequently encountered objects in hospitals and comprises 51,869 annotated objects. The dataset is characterized by dense annotation, with an average of 11.7 objects and 6.8 object categories per image. An object detection benchmark was established using the HIOD dataset and eight state-of-the-art object detectors. The benchmark provides a comprehensive evaluation of the performance of the selected object detectors on a large and diverse set of images of objects commonly seen in hospital environments. The results of the benchmark can be used to compare and analyze the performance of different object detectors and identify their strengths and weaknesses for use in hospital environments. In the benchmark, one-stage detectors have shown superior performance compared to two-stage detectors of similar parameter sizes. In particular, YOLOv6-L was able to attain a mean Average Precision (mAP) of 51.7% while operating at a detection speed of 255 FPS. The benchmark and dataset can serve as a valuable resource for researchers and practitioners in the field of computer vision and robotics, helping to advance the development of more effective and efficient object detection algorithms for developing automated operations in hospitals such as robotic disinfection and patient assistance.

Self-supervised semantic segmentation of retinal pigment epithelium cells in flatmount fluorescent microscopy images.

Morphological analyses with flatmount fluorescent images are essential to retinal pigment epithelial (RPE) aging studies and thus require accurate RPE cell segmentation. Although rapid technology advances in deep learning semantic segmentation have achieved great success in many biomedical research, the performance of these supervised learning methods for RPE cell segmentation is still limited by inadequate training data with high-quality annotations. To address this problem, we develop a Self-Supervised Semantic Segmentation (S4) method that utilizes a self-supervised learning strategy to train a semantic segmentation network with an encoder-decoder architecture. We employ a reconstruction and a pairwise representation loss to make the encoder extract structural information, while we create a morphology loss to produce the segmentation map. In addition, we develop a novel image augmentation algorithm (AugCut) to produce multiple views for self-supervised learning and enhance the network training performance. To validate the efficacy of our method, we applied our developed S4 method for RPE cell segmentation to a large set of flatmount fluorescent microscopy images, we compare our developed method for RPE cell segmentation with other state-of-the-art deep learning approaches. Compared with other state-of-the-art deep learning approaches, our method demonstrates better performance in both qualitative and quantitative evaluations, suggesting its promising potential to support large-scale cell morphological analyses in RPE aging investigations. The codes and the documentation are available at: https://github.com/jkonglab/S4_RPE.

Production of German Picture Postcards at the Western Front 1915–1916 as Exemplified by the Imagery of the Church Bell of Marquillies (Département du Nord, France)

During World War I the soldiers’ life at and behind the frontline was captured by personal photography. While heavily regulated and censored on the allied side, conditions were more relaxed for German soldiers, and for German officers, in particular. Drawing on a large sample set of images of the same subject matter, a French church bell with a patriotic, anti-German inscription, this paper surveys the private production of picture postcards by German soldiers. Initially photographed by a range of individuals, some images were eventually produced as printed lithographed postcards by regional German publishers. The processes and limitations of the personal versus commercial production of picture postcards are discussed.

Efficient and Accurate Hierarchical SfM Based on Adaptive Track Selection for Large-Scale Oblique Images

Image-based 3D modeling has been widely used in many areas. Structure from motion is the key to image-based reconstruction. However, the rapid growth of data poses challenges to current SfM solutions. A hierarchical SfM reconstruction methodology for large-scale oblique images is proposed. Firstly, match pairs are selected using positioning and orientation (POS) data and the terrain of the survey area. Then, images are divided to image groups by traversing the selected match pairs. After pairwise image matching, tracks are decimated using an adaptive track selection method. Thirdly, submaps are reconstructed from the image groups in parallel based on incremental SfM in the object space. A novel method based on statistics of the positional difference between common tracks is proposed to detect the outliers in submap merging. Finally, the reconstructed submaps are incrementally merged and optimized. The proposed methodology was used on a large oblique image set. The proposed methodology was compared with the state-of-the-art image-based reconstruction systems COLMAP and Metashape for SfM reconstruction. Experimental results show that the proposed methodology achieved the highest accuracy on the experimental dataset, i.e., about 22.37, and 3.52 times faster than COLMAP and Metashape, respectively. The experimental results demonstrate that the proposed hierarchical SfM methodology is accurate and efficient for large-scale oblique images.