Image Features Research Articles

HPIDN: A Hierarchical prior-guided iterative denoising network with global–local fusion for enhancing low-dose CT images

Low-dose computed tomography (LDCT) is an emerging medical diagnostic tool that reduces radiation exposure but suffers from noise retention. Current CNN-based LDCT denoising algorithms struggle to capture comprehensive global representations, impacting diagnostic accuracy. To address this, we propose a novel Hierarchical Prior-guided Iterative Denoising Network (HPIDN) for LDCT images, consisting of two main modules: the Dynamic Feature Extraction and Fusion Module (DFEFM) and the Feature-domain Iterative Denoising Module (FIDM). DFEFM dynamically captures a comprehensive representation, encompassing detailed local features in intra-relationships and complex global features in inter-relationships. It effectively guides the multi-stage iterative denoising process. FIDM hierarchically fuses the prior with image features from DFEFM by using the dual-domain attention fusion sub-network (DAFSN), enhancing denoising robustness and adaptability. This yields higher-quality images with reduced noise artifacts. Extensive experiments on the Mayo and ELCAP Datasets demonstrate the superior performance of our method quantitatively and qualitatively, improving diagnostic accuracy of lung diseases.

Fermented and Unfermented Cocoa Beans for Quality Identification Using Image Features

Fermented cocoa bean products are one of the high-quality requirements of the cocoa processing industry. On an automated industrial scale, early identification of cocoa bean quality is essential in the processing industry. This study aims to identify the condition of quality cocoa beans based on fermentation and non-fermentation characteristics. This study applies analysis based on static images taken using a camera with a distance variation of 5 cm, 10 cm, and 15 cm in both classes, with 500 image data each. The Feature extraction Approach uses the Oriented Gradient (HOG) method with a Support Vector Machine (SVM) classification technique. Image analysis of both object classes was also performed with a color change to show the dominance of the color pattern on the skin of the cocoa beans to be analyzed. The results showed that fermented cocoa beans show a color pattern and texture that tends to be darker and coarser than non-fermented cocoa beans. Computational results with performance analysis using Receiver Operating Characterisic (ROC) on both classes showed the results that the distance of 5 cm and 15 cm has 100% accuracy, but based on the best performance, comprehensively seen in terms of Precision, Recall, and F1-Score shows the best value is at a distance of 15 cm. The results of this research based on the literature review conducted have better achievements, thus enabling further research on the development of conveyor models with real-time video data for automation systems.

Hierarchical Graph Neural Network: A Lightweight Image Matching Model with Enhanced Message Passing of Local and Global Information in Hierarchical Graph Neural Networks

Graph Neural Networks (GNNs) have gained popularity in image matching methods, proving useful for various computer vision tasks like Structure from Motion (SfM) and 3D reconstruction. A well-known example is SuperGlue. Lightweight variants, such as LightGlue, have been developed with a focus on stacking fewer GNN layers compared to SuperGlue. This paper proposes the h-GNN, a lightweight image matching model, with improvements in the two processing modules, the GNN and matching modules. After image features are detected and described as keypoint nodes of a base graph, the GNN module, which primarily aims at increasing the h-GNN’s depth, creates successive hierarchies of compressed-size graphs from the base graph through a clustering technique termed SC+PCA. SC+PCA combines Principal Component Analysis (PCA) with Spectral Clustering (SC) to enrich nodes with local and global information during graph clustering. A dual non-contrastive clustering loss is used to optimize graph clustering. Additionally, four message-passing mechanisms have been proposed to only update node representations within a graph cluster at the same hierarchical level or to update node representations across graph clusters at different hierarchical levels. The matching module performs iterative pairwise matching on the enriched node representations to obtain a scoring matrix. This matrix comprises scores indicating potential correct matches between the image keypoint nodes. The score matrix is refined with a ‘dustbin’ to further suppress unmatched features. There is a reprojection loss used to optimize keypoint match positions. The Sinkhorn algorithm generates a final partial assignment from the refined score matrix. Experimental results demonstrate the performance of the proposed h-GNN against competing state-of-the-art (SOTA) GNN-based methods on several image matching tasks under homography, estimation, indoor and outdoor camera pose estimation, and 3D reconstruction on multiple datasets. Experiments also demonstrate improved computational memory and runtime, approximately 38.1% and 26.14% lower than SuperGlue, and an average of about 6.8% and 7.1% lower than LightGlue. Future research will explore the effects of integrating more recent simplicial message-passing mechanisms, which concurrently update both node and edge representations, into our proposed model.

CNN Algorithm with SIFT to Enhance the Arabic Sign Language Recognition

Sign language is used as a primary means of communication by millions of people who suffer from hearing problems. The unhearing people used visual language to interact with each other, Represented in sign language. There are features that the hearing impaired use to understand each other, which are difficult for normal people to understand. Therefore, deaf people will struggle to interact with society. This research aims to introduce a system for recognizing hand gestures in Arabic Sign Language (ArSL) through training the Convolutional Neural Network (CNN) on the images of ArSL gestures launched by the University of Prince Mohammad Bin Fahd, Saudi Arabia. A Scale Invariant Feature Transform (SIFT) algorithm is used for creating the feature vectors that contain shape, finger position, size, center points of palm, and hand margin by extracting the Important features for images of ArSL and transforming them to points of the vector. The accuracy of the proposed system is 97% using the SIFT with CNN, and equal to 94.8% nearly without SIFT. Finally, the proposed system was tried and tested on a group of persons and its effectiveness was proven after considering their observations.

A Hybrid Approach for CT Image Noise Reduction Combining Method Noise-CNN and Shearlet Transform

The presence of gaussian noise commonly weakens the diagnostic precision of low-dose CT imaging. A novel CT image denoising technique that integrates the non-subsampled shearlet transform (NSST) with Bayesian thresholding, and incorporates a modern method noise Deep Convolutional neural network (DCNN) based post-processing operation on denoised images to strengthen low-dose CT imaging quality. The hybrid method commences with NSST and Bayesian thresholding to mitigate the initial noise while preserving crucial image features, such as corners and edges. The novel aspect of the proposed approach is its successive application of a DnCNN on initial denoised image, which learns and removes residual noise patterns from denoised images, thereby enhancing fine detail preservation. This dual-phase strategy addresses both noise suppression and image-detail preservation. The proposed technique is evaluated through the use of metrics, such as PSNR, SNR, SSIM, ED, and UIQI. The results demonstrate that the hybrid approach outperforms standard denoising techniques in preserving image quality and fine details.

Multi-Modal Fusion Network with Multi-Head Self-Attention for Injection Training Evaluation in Medical Education

The COVID-19 pandemic has significantly disrupted traditional medical training, particularly in critical areas such as the injection process, which require expert supervision. To address the challenges posed by reduced face-to-face interactions, this study introduces a multi-modal fusion network designed to evaluate the timing and motion aspects of the injection training process in medical education. The proposed framework integrates 3D reconstructed data and 2D images of hand movements during the injection process. The 3D data are preprocessed and encoded by a Long Short-Term Memory (LSTM) network to extract temporal features, while a Convolutional Neural Network (CNN) processes the 2D images to capture detailed image features. These encoded features are then fused and refined through a proposed multi-head self-attention module, which enhances the model’s ability to capture and weigh important temporal and image dynamics in the injection process. The final classification of the injection process is conducted by a classifier module. The model’s performance was rigorously evaluated using video data from 255 subjects with assessments made by professional physicians according to the Objective Structured Assessment of Technical Skill—Global Rating Score (OSATS-GRS)[B] criteria for time and motion evaluation. The experimental results demonstrate that the proposed data fusion model achieves an accuracy of 0.7238, an F1-score of 0.7060, a precision of 0.7339, a recall of 0.7238, and an AUC of 0.8343. These findings highlight the model’s potential as an effective tool for providing objective feedback in medical injection training, offering a scalable solution for the post-pandemic evolution of medical education.

SSResNeXt: A Novel Deep Learning Architecture for Multi-class Breast Cancer Pathological Image Classification

Multi-class classification of breast cancer pathological images remains challenging due to complex image features and limited datasets. This study proposes SSResNeXt, a novel deep learning architecture incorporating a new Small-SE-ResNeXt Block with asymmetric convolutions and channel attention mechanisms. Evaluated on the BreaKHis dataset, SSResNeXt achieves state-of-the-art performance with accuracies of 95.2%, 94.0%, 92.7%, and 93.5% for 40X, 100X, 200X, and 400X magnification scales, respectively. Comparative experiments demonstrate SSResNeXt's superior performance over existing models, including ResNet and Swin Transformer variants. The proposed architecture offers improved feature extraction capabilities for breast cancer pathological images without significantly increasing model complexity, providing a promising tool for computer-aided diagnosis systems.

Low-Light Image Enhancement Network Using Informative Feature Stretch and Attention

Low-light images often exhibit reduced brightness, weak contrast, and color distortion. Consequently, enhancing low-light images is essential to make them suitable for computer vision tasks. Nevertheless, addressing this task is particularly challenging because of the inherent constraints posed by low-light environments. In this study, we propose a novel low-light image enhancement network using adaptive feature stretching and informative attention. The proposed network architecture mainly includes an adaptive feature stretch block designed to extend the narrow range of image features to a broader range. To achieve improved image restoration, an informative attention block is introduced to assign weight to the output features from the adaptive feature stretch block. We conduct comprehensive experiments on widely used benchmark datasets to assess the effectiveness of the proposed network. The experimental results show that the proposed low-light image enhancement network yields satisfactory results compared with existing state-of-the-art methods from both subjective and objective perspectives while maintaining acceptable network complexity.

Assessment of fish health status for waterless transportation based on image features and deep learning models

In the current aquaculture industry, the detection of physiological stress indicators in fish during waterless transportation is considered to be an effective method to monitor their health status. However, commonly used detection methods are invasive and interfere with the accuracy of detection results. In order to solve this problem, this study takes grouper as an example, combines multi-sensor physiological stress indicators monitoring, and designs a health status assessment and classification model based on image features and deep learning models, which inputs real-time fish image feature information data acquired during low-temperature waterless transportation into the optimised CNN-BiGRU classification model based on whale optimisation algorithm (WOA) proposed in this study, so as to achieve an accurate assessment of the health status of the fish. By inputting the image feature datasets based on time series and classified into the WOA-CNN-BiGRU, BiGRU, and GRU classification models for training, validation, and testing, respectively, the effectiveness of the classification performance of the models was verified by using accuracy, precision, recall and F1-value composite scores as the metrics. It was found that the WOA-CNN-BiGRU model had excellent classification performance, followed by the BiGRU and GRU models, especially in classifying the health status assessment of small size grouper. This study proposes an efficient and inexpensive method for real-time monitoring and accurate assessment of the health status of fish during waterless transportation, and also provides a reference for vigour monitoring and health assessment of other similar aquatic products.

Estimation of Cotton SPAD Based on Multi-Source Feature Fusion and Voting Regression Ensemble Learning in Intercropping Pattern of Cotton and Soybean

The accurate estimation of soil plant analytical development (SPAD) values in cotton under various intercropping patterns with soybean is crucial for monitoring cotton growth and determining a suitable intercropping pattern. In this study, we utilized an unmanned aerial vehicle (UAV) to capture visible (RGB) and multispectral (MS) data of cotton at the bud stage, early flowering stage, and full flowering stage in a cotton–soybean intercropping pattern in the Yellow River Delta region of China, and we used SPAD502 Plus and tapeline to collect SPAD and cotton plant height (CH) data of the cotton canopy, respectively. We analyzed the differences in cotton SPAD and CH under different intercropping ratio patterns. It was conducted using Pearson correlation analysis between the RGB features, MS features, and cotton SPAD, then the recursive feature elimination (RFE) method was employed to select image features. Seven feature sets including MS features (five vegetation indices + five texture features), RGB features (five vegetation indices + cotton cover), and CH, as well as combinations of these three types of features with each other, were established. Voting regression (VR) ensemble learning was proposed for estimating cotton SPAD and compared with the performances of three models: random forest regression (RFR), gradient boosting regression (GBR), and support vector regression (SVR). The optimal model was then used to estimate and visualize cotton SPAD under different intercropping patterns. The results were as follows: (1) There was little difference in the mean value of SPAD or CH under different intercropping patterns; a significant positive correlation existed between CH and SPAD throughout the entire growth period. (2) All VR models were optimal when each of the seven feature sets were used as input. When the features set was MS + RGB, the determination coefficient (R2) of the validation set of the VR model was 0.902, the root mean square error (RMSE) was 1.599, and the relative prediction deviation (RPD) was 3.24. (3) When the features set was CH + MS + RGB, the accuracy of the VR model was further improved, compared with the feature set MS + RGB, the R2 and RPD were increased by 1.55% and 8.95%, respectively, and the RMSE was decreased by 7.38%. (4) In the intercropping of cotton and soybean, cotton growing under 4:6 planting patterns was better. The results can provide a reference for the selection of intercropping patterns and the estimation of cotton SPAD.

Relationship between EZH2 expression and prognosis of patients with hepatocellular carcinoma using a pathomics predictive model

BackgroundEnhancer of zeste 2 polycomb repressive complex 2 subunit (EZH2) is overexpressed in hepatocellular carcinoma, promoting tumorigenesis and correlating with poor prognosis. Traditional histopathological examinations are insufficient to accurately predict hepatocellular carcinoma (HCC) survival; however, pathomics models can predict EZH2 expression and HCC prognosis. This study aimed to investigate the relationship between pathomics features and EZH2 expression for predicting overall survival of patients with HCC. MethodsWe analyzed 267 patients with HCC from the Cancer Genome Atlas database, with available pathological images and gene expression data. RNA sequencing data were divided into high and low EZH2 expression groups for prognosis and survival analysis. Pathological image features were screened using mRMR_RFE. A pathological model was constructed using a gradient boosting machine (GBM) algorithm, and efficiency evaluation and survival analysis of the model were performed. The R package “survminer” took the pathomics score (PS) cutoff value of 0.4628 to divide the patients into two groups: high and low PS expression. Survival analyses included Kaplan–Meier curve analysis, univariate and multivariate Cox regression analyses, and interaction tests. Potential pathomechanisms were explored through enrichment, differential, immune cell infiltration abundance, and gene mutation analyses. ResultEZH2 was highly expressed in tumor samples but poorly expressed in normal tissue samples. Univariate and multivariate Cox regression analyses revealed that EZH2 was an independent risk factor for HCC (hazard ratio [HR], 2.792 and 3.042, respectively). Seven imaging features were selected to construct a pathomics model to predict EZH2. Decision curve analysis showed that the model had high clinical utility. Multivariate Cox regression analysis showed that high PS expression was an independent risk factor for HCC prognosis (HR, 2.446). The Kaplan–Meier curve showed that high PS expression was a risk factor for overall survival. ConclusionEZH2 expression can affect the prognosis of patients with liver cancer. Our pathological model could predict EZH2 expression and prognosis of patients with HCC with high accuracy and robustness, making it a new and potentially valuable tool.

Computed tomography-based multi-organ radiomics nomogram model for predicting the risk of esophagogastric variceal bleeding in cirrhosis.

Radiomics has been used in the diagnosis of cirrhosis and prediction of its associated complications. However, most current studies predict the risk of esophageal variceal bleeding (EVB) based on image features at a single level, which results in incomplete data. Few studies have explored the use of global multi-organ radiomics for non-invasive prediction of EVB secondary to cirrhosis. To develop a model based on clinical and multi-organ radiomic features to predict the risk of first-instance secondary EVB in patients with cirrhosis. In this study, 208 patients with cirrhosis were retrospectively evaluated and randomly split into training (n = 145) and validation (n = 63) cohorts. Three areas were chosen as regions of interest for extraction of multi-organ radiomic features: The whole liver, whole spleen, and lower esophagus-gastric fundus region. In the training cohort, radiomic score (Rad-score) was created by screening radiomic features using the inter-observer and intra-observer correlation coefficients and the least absolute shrinkage and selection operator method. Independent clinical risk factors were selected using multivariate logistic regression analyses. The radiomic features and clinical risk variables were combined to create a new radiomics-clinical model (RC model). The established models were validated using the validation cohort. The RC model yielded the best predictive performance and accurately predicted the EVB risk of patients with cirrhosis. Ascites, portal vein thrombosis, and plasma prothrombin time were identified as independent clinical risk factors. The area under the receiver operating characteristic curve (AUC) values for the RC model, Rad-score (liver + spleen + esophagus), Rad-score (liver), Rad-score (spleen), Rad-score (esophagus), and clinical model in the training cohort were 0.951, 0.930, 0.801, 0.831, 0.864, and 0.727, respectively. The corresponding AUC values in the validation cohort were 0.930, 0.886, 0.763, 0.792, 0.857, and 0.692. In patients with cirrhosis, combined multi-organ radiomics and clinical model can be used to non-invasively predict the probability of the first secondary EVB.

Enhanced gastric cancer classification and quantification interpretable framework using digital histopathology images

Recent developments have highlighted the critical role that computer-aided diagnosis (CAD) systems play in analyzing whole-slide digital histopathology images for detecting gastric cancer (GC). We present a novel framework for gastric histology classification and segmentation (GHCS) that offers modest yet meaningful improvements over existing CAD models for GC classification and segmentation. Our methodology achieves marginal improvements over conventional deep learning (DL) and machine learning (ML) models by adaptively focusing on pertinent characteristics of images. This contributes significantly to our study, highlighting that the proposed model, which performs well on normalized images, is robust in certain respects, particularly in handling variability and generalizing to different datasets. We anticipate that this robustness will lead to better results across various datasets. An expectation-maximizing Naïve Bayes classifier that uses an updated Gaussian Mixture Model is at the heart of the suggested GHCS framework. The effectiveness of our classifier is demonstrated by experimental validation on two publicly available datasets, which produced exceptional classification accuracies of 98.87% and 97.28% on validation sets and 98.47% and 97.31% on test sets. Our framework shows a slight but consistent improvement over previously existing techniques in gastric histopathology image classification tasks, as demonstrated by comparative analysis. This may be attributed to its ability to capture critical features of gastric histopathology images better. Furthermore, using an improved Fuzzy c-means method, our study produces good results in GC histopathology picture segmentation, outperforming state-of-the-art segmentation models with a Dice coefficient of 65.21% and a Jaccard index of 60.24%. The model’s interpretability is complemented by Grad-CAM visualizations, which help understand the decision-making process and increase the model’s trustworthiness for end-users, especially clinicians.

Iterative Optimization-Enhanced Contrastive Learning for Multimodal Change Detection

Multimodal change detection (MCD) harnesses multi-source remote sensing data to identify surface changes, thereby presenting prospects for applications within disaster management and environmental surveillance. Nonetheless, disparities in imaging mechanisms across various modalities impede the direct comparison of multimodal images. In response, numerous methodologies employing deep learning features have emerged to derive comparable features from such images. Nevertheless, several of these approaches depend on manually labeled samples, which are resource-intensive, and their accuracy in distinguishing changed and unchanged regions is not satisfactory. In addressing these challenges, a new MCD method based on iterative optimization-enhanced contrastive learning is proposed in this paper. With the participation of positive and negative samples in contrastive learning, the deep feature extraction network focuses on extracting the initial deep features of multimodal images. The common projection layer unifies the deep features of two images into the same feature space. Then, the iterative optimization module expands the differences between changed and unchanged areas, enhancing the quality of the deep features. The final change map is derived from the similarity measurements of these optimized features. Experiments conducted across four real-world multimodal datasets, benchmarked against eight well-established methodologies, incontrovertibly illustrate the superiority of our proposed approach.

Overview of Structure from Motion

In this paper is given an overview of tools and techniques for obtain information about the geometry of 3D scenes from 2D images. The generating three-dimensional structure from a series of 2D images or video of scenes is known as Structure from Motion (SfM). A central tenet of structure from motion is that, given the position of a feature in one image, it is possible to find the corresponding position of the same feature in successive images. We described methods for the simultaneous recovery of 3D points and camera projection matrices using corresponding image points in multiple views. Structure from Motion (SfM) is a fascinating field within computer vision that seeks to reconstruct a three-dimensional structure of the environment from a sequence of two-dimensional images. At the heart of SfM lies a set of sophisticated algorithms that enable the extraction of spatial information from a series of images. Two public repositories that may be of use are Open SfM and Colmap.

DINO-Mix enhancing visual place recognition with foundational vision model and feature mixing

Using visual place recognition (VPR) technology to ascertain the geographical location of publicly available images is a pressing issue. Although most current VPR methods achieve favorable results under ideal conditions, their performance in complex environments, characterized by lighting variations, seasonal changes, and occlusions, is generally unsatisfactory. Therefore, obtaining efficient and robust image feature descriptors in complex environments is a pressing issue. In this study, we utilized the DINOv2 model as the backbone for trimming and fine-tuning to extract robust image features and employed a feature mix module to aggregate image features, resulting in globally robust and generalizable descriptors that enable high-precision VPR. We experimentally demonstrated that the proposed DINO-Mix outperforms the current state-of-the-art (SOTA) methods. Using test sets having lighting variations, seasonal changes, and occlusions such as Tokyo24/7, Nordland, and SF-XL-Testv1, our proposed architecture achieved Top-1 accuracy rates of 91.75%, 80.18%, and 82%, respectively, and exhibited an average accuracy improvement of 5.14%. In addition, we compared it with other SOTA methods using representative image retrieval case studies, and our architecture outperformed its competitors in terms of VPR performance. Furthermore, we visualized the attention maps of DINO-Mix and other methods to provide a more intuitive understanding of their respective strengths. These visualizations serve as compelling evidence of the superiority of the DINO-Mix framework in this domain.

VQGNet: An Unsupervised Defect Detection Approach for Complex Textured Steel Surfaces

Defect detection on steel surfaces with complex textures is a critical and challenging task in the industry. The limited number of defect samples and the complexity of the annotation process pose significant challenges. Moreover, performing defect segmentation based on accurate identification further increases the task’s difficulty. To address this issue, we propose VQGNet, an unsupervised algorithm that can precisely recognize and segment defects simultaneously. A feature fusion method based on aggregated attention and a classification-aided module is proposed to segment defects by integrating different features in the original images and the anomaly maps, which direct the attention to the anomalous information instead of the irregular complex texture. The anomaly maps are generated more confidently using strategies for multi-scale feature fusion and neighbor feature aggregation. Moreover, an anomaly generation method suitable for grayscale images is introduced to facilitate the model’s learning on the anomalous samples. The refined anomaly maps and fused features are both input into the classification-aided module for the final classification and segmentation. VQGNet achieves state-of-the-art (SOTA) performance on the industrial steel dataset, with an I-AUROC of 99.6%, I-F1 of 98.8%, P-AUROC of 97.0%, and P-F1 of 80.3%. Additionally, ViT-Query demonstrates robust generalization capabilities in generating anomaly maps based on the Kolektor Surface-Defect Dataset 2.

Research on rapid prediction method of laser cladding deposited layer state based on molten pool texture sequence

A texture sequence feature combining texture statistical characteristics with time series is proposed for rapid prediction of the state of laser cladding deposited layer. Texture features in the molten pool images are extracted using high-order dense texture descriptor operators and transformed into probability sequences with the aid of multi-cell histograms. Texture features are screened based on the types of points of interest to improve the sequence, enhancing the prediction accuracy and speed for deposited layer. For different types of points of interest, the optimal texture sequence obtained achieves an improvement of more than 6% in prediction accuracy for deposited layer state compared to the full-texture sequence. The proposed texture sequence features, when compared with molten pool image features, exhibit an average improvement in prediction speed of more than 17 times across different deep learning models. Overall, the results indicate that this method significantly improves prediction speed while maintaining accuracy, satisfying the requirements of high-speed monitoring in laser cladding.

A Transformer-Based Image-Guided Depth-Completion Model with Dual-Attention Fusion Module.

Depth information is crucial for perceiving three-dimensional scenes. However, depth maps captured directly by depth sensors are often incomplete and noisy, our objective in the depth-completion task is to generate dense and accurate depth maps from sparse depth inputs by fusing guidance information from corresponding color images obtained from camera sensors. To address these challenges, we introduce transformer models, which have shown great promise in the field of vision, into the task of image-guided depth completion. By leveraging the self-attention mechanism, we propose a novel network architecture that effectively meets these requirements of high accuracy and resolution in depth data. To be more specific, we design a dual-branch model with a transformer-based encoder that serializes image features into tokens step by step and extracts multi-scale pyramid features suitable for pixel-wise dense prediction tasks. Additionally, we incorporate a dual-attention fusion module to enhance the fusion between the two branches. This module combines convolution-based spatial and channel-attention mechanisms, which are adept at capturing local information, with cross-attention mechanisms that excel at capturing long-distance relationships. Our model achieves state-of-the-art performance on both the NYUv2 depth and SUN-RGBD depth datasets. Additionally, our ablation studies confirm the effectiveness of the designed modules.

Spatio-temporal fusion with motion masks for the moving small target detection from remote-sensing videos

Moving small target detection is an eye-catching research topic, full of great challenges in object detection field. Its primary difficulties exist in (i) the very small size proportion of the target region in background images, (ii) the quite weak contrast to background and (iii) the perception of slight target motion. Currently-existing detection methods mainly utilize the image features from spatial domain. This kind of features is quite weak to small targets. More new feature kinds need to be concerned. In view of these, besides traditional image clues, motion features are currently attracting more and more attention in small target detection. In order to promote detection performance, this paper proposes a Spatio-Temporal Fusion Network (STFNet) with two parallel feature extraction branches for detecting moving small targets. In this network, one branch is designed to capture the traditional semantic features from spatial domain, and the other is devised to perceive the slight target motion features hidden in temporal domain. Meanwhile, to optimize the extraction of spatio-temporal features, a group of motion masks is specially designed to guide feature extractors to pay more precise attention to the positions where small targets appear. Moreover, we design a new bridging module to enhance the cross-domain and cross-scale fusion of spatio-temporal features. Through extensive comparison and ablation experiments on seven sub-datasets, it is proved that our new STFNet is effective and it is obviously superior to the compared ones in detecting moving small targets from satellite sequence images. It achieves a Precision of 0.93, an mAP50 of 71.10% and an F1 of 0.84, evidently higher than existing state-of-the-art methods. Our codes are available at https://github.com/UESTC-nnLab/STF.