High-level Semantic Information Research Articles

Spatial Focus Attention for Fine-Grained Skeleton-Based Action Tasks

Dynamic skeletal data has been widely studied for human action tasks due to its high-level semantic information and less data than RGB features. However, attention-based previous methods fail to focus on the local grouped joint dependence of the human body, which is vital to distinguishing various actions in fine-grained tasks, such as skeletal action segmentation and recognition. This work proposes spatial focus attention for the fine-grained skeleton-based action tasks. Specifically, we decouple the attention map to enhance the grouped joint dependence adaptively by the decouple probability. To further focus on local grouped dependence, the tree structural attention maps can be built by hierarchical decoupling and guide the model to focus on complementary local dependence in the different leaf nodes. Our proposed approach achieves state-of-the-art performance on fine-grained skeleton-based human action segmentation tasks (MCFS-22) and recognition tasks (FSD-10). Besides, on the coarse-grained dataset (NTU-60), the proposed spatial focus attention also achieves outstanding performance.

Upsampling Autoencoder for Self-Supervised Point Cloud Learning

In computer-aided design (CAD) community, the point cloud data is pervasively applied in reverse engineering, where the point cloud analysis plays an important role. While a large number of supervised learning methods have been proposed to handle the unordered point clouds and demonstrated their remarkable success, their performance and applicability are limited to the costly data annotation. In this work, we propose a novel self-supervised pretraining model for point cloud learning without human annotations, which relies solely on upsampling operation to perform feature learning of point cloud in an effective manner. The key premise of our approach is that upsampling operation encourages the network to capture both high-level semantic information and low-level geometric information of the point cloud, thus the downstream tasks such as classification and segmentation will benefit from the pre-trained model. Specifically, our method first conducts the random subsampling from the input point cloud at a low proportion e.g., 12.5%. Then, we feed them into an encoder-decoder architecture, where an encoder is devised to operate only on the subsampled points, along with a upsampling decoder is adopted to reconstruct the original point cloud based on the learned features. Finally, we design a novel joint loss function which enforces the upsampled points to be similar with the original point cloud and uniformly distributed on the underlying shape surface. By adopting the pre-trained encoder weights as initialisation of models for downstream tasks, we find that our UAE outperforms previous state-of-the-art methods in shape classification, part segmentation and point cloud upsampling tasks. Code will be made publicly available upon acceptance.

An Algorithm for Target Detection of Engineering Vehicles Based on Improved CenterNet

Aiming at the problems of low target image resolution, insufficient target feature extraction, low detection accuracy and poor real time in remote engineering vehicle detection, an improved CenterNet target detection model is proposed in this paper. Firstly, EfficientNet-B0 with Efficient Channel Attention (ECA) module is used as the basic network, which increases the quality and speed of feature extraction and reduces the number of model parameters. Then, the proposed Adaptive Fusion Bidirectional Feature Pyramid Network (AF-BiFPN) module is applied to fuse the features of different feature layers. Furthermore, the feature information of engineering vehicle targets is added by making full use of the high-level semantic and low-level fine-grained feature information of the target, which overcomes the problem that the original CenterNet network did not perform well in small target detection and improve the detection accuracy of the network. Finally, the tag coding strategy and bounding box regression method of CenterNet are optimized by introducing positioning quality loss. The accuracy of target prediction is increased by joint prediction of center position and target size. Experimental results show that the mean Average Precision (mAP) of the improved CenterNet model is 94.74% on the engineering vehicle dataset, and the detection rate is 29 FPS. Compared with the original CenterNet model based on ResNet-18, the detection accuracy of this model is improved by 16.29%, the detection speed is increased by 9 FPS, and the memory usage is reduced by 43 MB. Compared with YOLOv3 and YOLOv4, the mAP of this model is improved by 19.9% and 5.61% respectively. The proposed method can detect engineering vehicles more quickly and accurately in far distance. It has obvious advantages in target detection compared with traditional methods.

Small target detection method based on feature fusion for deep learning in state grid environment evaluation

Aiming at the problem that small and medium-sized targets cannot be detected in real time in high-resolution images, a new target detection network model is proposed. Firstly, the residual network RESNET is used as the basic network structure, an additional pyramid network model is added, and the pool layer is used to increase the number of hierarchical feature mapping. Then, the feature map is deconvoluted, and the high-level semantic feature map information and shallow feature map information are fused. Finally, the target is detected. Based on the analysis of the experimental results, compared with the existing target detection network model, the deep learning network model using feature fusion techniques has a detection accuracy of 80.2% on the standard dataset Pascal voc2007, and the detection speed reaches 27 frames per second, which meet the requirements of high-resolution image real-time monitoring and small target detection.

Image fusion in the loop of high-level vision tasks: A semantic-aware real-time infrared and visible image fusion network

Infrared and visible image fusion aims to synthesize a single fused image that not only contains salient targets and abundant texture details but also facilitates high-level vision tasks. However, the existing fusion algorithms unilaterally focus on the visual quality and statistical metrics of fused images but ignore the demands of high-level vision tasks. To address these challenges, this paper bridges the gap between image fusion and high-level vision tasks and proposes a semantic-aware real-time image fusion network (SeAFusion). On the one hand, we cascade the image fusion module and semantic segmentation module and leverage the semantic loss to guide high-level semantic information to flow back to the image fusion module, which effectively boosts the performance of high-level vision tasks on fused images. On the other hand, we design a gradient residual dense block (GRDB) to enhance the description ability of the fusion network for fine-grained spatial details. Extensive comparative and generalization experiments demonstrate the superiority of our SeAFusion over state-of-the-art alternatives in terms of maintaining pixel intensity distribution and preserving texture detail. More importantly, the performance comparison of various fusion algorithms in task-driven evaluation reveals the natural advantages of our framework in facilitating high-level vision tasks. In addition, the superior running efficiency allows our algorithm to be effortlessly deployed as a real-time pre-processing module for high-level vision tasks. The source code will be released at https://github.com/Linfeng-Tang/SeAFusion.

Real-Time One-Stream Semantic-Guided Refinement Network for RGB-Thermal Salient Object Detection

Salient Object Detection (SOD) has been widely used in practical applications such as multi-sensor image fusion, remote sensing, and defect detection. Recently, SOD from RGB and Thermal (T) has been rapidly developed due to its robustness to extreme situations like low illumination and occlusion. However, existing methods all utilize a dual-stream encoder, which significantly increases the computation burdens and hinders real-world deployment. To this end, we propose a real-time One-stream Semantic-guided Refinement Network (OSRNet) for RGB-T SOD. Specifically, we firstly fuse the RGB and T via concatenation, addition, and multiplication operations to dig the complementary information between each modality. The efficient early fusion not only facilitates the information exchange between each modality but also avoids the cumbersome dual-stream encoder structure. Then, the light-weight decoder is proposed, making the high-level semantic information filter the low-level noisy features and gradually refine the final prediction. Also, we apply deep supervision to make the training procedure more stable and fast. Due to the early fusion strategy, OSRNet can run at a real-time speed (53-60<i>fps</i>) on a single GPU. Extensive quantitative and qualitative experiments show our network outperforms eleven state-of-the-art methods in terms of seven evaluation metrics. Our codes have been released at: https://github.com/huofushuo/OSRNet.

Detect and Locate: Exposing Face Manipulation by Semantic- and Noise-Level Telltales

The technological advancements of deep learning have enabled sophisticated face manipulation schemes, raising severe trust issues and security concerns in modern society. Generally speaking, detecting manipulated faces and locating the potentially altered regions are challenging tasks. Herein, we propose a conceptually simple but effective method to efficiently detect forged faces in an image while simultaneously locating the manipulated regions. The proposed scheme relies on a segmentation map that delivers meaningful high-level semantic information clues about the image. Furthermore, a noise map is estimated, playing a complementary role in capturing low-level clues and subsequently empowering decision-making. Finally, the features from these two modules are combined to distinguish fake faces. Extensive experiments show that the proposed model achieves state-of-the-art detection accuracy and remarkable localization performance.

SPNet: Spectral Patching End-to-End Classification Network for UAV-Borne Hyperspectral Imagery With High Spatial and Spectral Resolutions

In deep learning (DL)-based hyperspectral imagery classification, &#x201C;spatial patching&#x201D; is primarily used as a preprocessing for incorporating local spatial information. This operation can help to promote classification accuracy but it is facing new challenges in the unmanned aerial vehicle (UAV)-borne hyperspectral imagery with high spatial and spectral resolutions (H² imagery). The ground objects&#x2019; various spatial scales result in it being challenging to determine the optimal size for the spatial patches. In addition, due to the severe spectral variability and spatial heterogeneity of the H² imagery, &#x201C;spatial patching&#x201D; only exploits the local spatial information and results in serious salt-and-pepper (SP) noise and isolated areas in the classification maps. In this article, to address these issues, a novel spectral patching network (SPNet) with an end-to-end DL architecture is proposed for UAV-borne H² imagery classification. The &#x201C;spectral patching&#x201D; approach is proposed to preserve the global spatial information and almost all the spectral information of the original hyperspectral imagery. An end-to-end deep encoder&#x2013;decoder network is then constructed based on the spectral patching mechanism, which introduces the deep residual network (ResNet) and atrous spatial pyramid pooling (ASPP) modules to extract multiscale high-level semantic information for the H² imagery classification. The experimental results obtained with the Wuhan UAV-borne H² imagery (WHU-Hi) UAV-borne hyperspectral data set demonstrate that SPNet can achieve state-of-the-art accuracy and visualization performance in the classification of H² imagery.

WH-MAVS: A Novel Dataset and Deep Learning Benchmark for Multiple Land Use and Land Cover Applications

Over the past decade, many excellent data sharing efforts have enriched the remote sensing scene classification (SC) methods. These datasets have achieved great success in complex high-level semantic information interpretation. However, most existing datasets are collected from standard and ungeoreferenced image patches for algorithm training and evaluation. These datasets do not fit for practical applications and cannot be directly applied in further geographical study. Accordingly, we provide a large range high-resolution SC dataset with multiple time phases, called “ <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">W</b> u <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">h</b> an <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">M</b> ulti <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">a</b> pplication <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">V</b> HR <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">S</b> cene classification dataset (WH-MAVS).” It facilitates the study of SC and scene change detection (SCD) algorithms. Moreover, it can also be directly employed to perform a variety of real-life land use application tasks. To the best of our knowledge, this is the first free, publicly available, georeferenced, and annotated dataset to cover almost an entire megacity. The WH-MAVS was collected and annotated from Google Earth imagery with the same spatial resolution and uniform nonoverlapping patch size, covering the central area of Wuhan, China. The total number of scene samples is 47 137, which belong to 14 classes with 23 567 labeled patches for each time phase in 2014 and 2016, respectively. The geographic coordinates of all samples in both time phases exhibit one-to-one correspondence with 23 202 unchanged image patches of scene categories and 365 changed ones. The distribution of the number of samples in each class is highly imbalanced; moreover, there are large intraclass differences and indistinguishable interclass variances. These characteristics are closer to the real land use/land cover application tasks and introduce further challenges to the related algorithm research. In addition, we conducted benchmark experiments on SC and SCD based on the WH-MAVS dataset with widely used deep learning models. DenseNet169 was found to achieve the best performance. The overall accuracies are 91.07% and 92.09%, respectively, in the 2014 and 2016 validation sets of WH-MAVS. Furthermore, SCD obtained by DenseNet169 has a binary change detection accuracy of 89.56% and a multiple (from–to) change detection accuracy of 86.70%. Over and above the research value of the algorithm, it is also proven to have practical applications in fields such as urban planning, landscape pattern analysis, and urban dynamic monitoring and analysis.

Cross-Modality Image Matching Network With Modality-Invariant Feature Representation for Airborne-Ground Thermal Infrared and Visible Datasets

Thermal infrared (TIR) remote-sensing imagery can allow objects to be imaged clearly at night through the long-wave infrared, so that the fusion of thermal infrared and visible (VIS) imagery is a way to improve the remote-sensing interpretation ability. However, due to the large radiation difference between the two kinds of images, it is very difficult to match them. One of the most important issues is the lack of comprehensive consideration of the modality-specific information and modality-shared information, which makes it difficult for the existing methods to obtain a modality-invariant feature representation. In this article, a cross-modality image matching network, which we refer to as CMM-Net, is proposed to realize thermal infrared and visible image matching by learning a modality-invariant feature representation. First, in order to extract the modality-specific features of the imagery, the framework constructs a shallow two-branch network to make full use of the modality-specific information, without sharing parameters. Second, in order to extract the high-level semantic information between the different modalities, modality-shared layers are embedded into the deep layers of the network. In addition, three novel loss functions are designed and combined to learn the modality-invariant feature representation, that is, the discriminative loss of the non-corresponding features in the same modality, the cross-modality loss of the corresponding features between different modalities, and the cross-modality triplet (CMT) loss. The multimodal matching experiments conducted with ground- and airborne-based thermal infrared images and visible images showed that the proposed method outperforms the existing image matching methods by about 2% and 6% for the ground and airborne images, respectively.

Residual based attention-Unet combing DAC and RMP modules for automatic liver tumor segmentation in CT.

Purpose: Due to the complex distribution of liver tumors in the abdomen, the accuracy of liver tumor segmentation cannot meet the needs of clinical assistance yet. This paper aims to propose a new end-to-end network to improve the segmentation accuracy of liver tumors from CT. Method: We proposed a hybrid network, leveraging the residual block, the context encoder (CE), and the Attention-Unet, called ResCEAttUnet. The CE comprises a dense atrous convolution (DAC) module and a residual multi-kernel pooling (RMP) module. The DAC module ensures the network derives high-level semantic information and minimizes detailed information loss. The RMP module improves the ability of the network to extract multi-scale features. Moreover, a hybrid loss function based on cross-entropy and Tversky loss function is employed to distribute the weights of the two-loss parts through training iterations. Results: We evaluated the proposed method in LiTS17 and 3DIRCADb databases. It significantly improved the segmentation accuracy compared to state-of-the-art methods. Conclusions: Experimental results demonstrate the satisfying effects of the proposed method through both quantitative and qualitative analyses, thus proving a promising tool in liver tumor segmentation.

A novel motion recognition method based on improved two-stream convolutional neural network and sparse feature fusion

Motion recognition is a hot topic in the field of computer vision. It is a challenging task. Motion recognition analysis is closely related to the network input, network structure and feature fusion. Due to the noise in the video, traditional methods cannot better obtain the feature information resulting in the problem of inaccurate motion recognition. Feature selection directly affects the efficiency of recognition, and there are still many problems to be solved in the multi-level feature fusion process. In this paper, we propose a novel motion recognition method based on an improved two-stream convolutional neural network and sparse feature fusion. In the low-rank space, because sparse features can effectively capture the information of motion objects in the video, meanwhile, we supplement the network input data, in view of the lack of information interaction in the network, we fuse the high-level semantic information and low-level detail information to recognize the motions by introducing attention mechanism, which makes the performance of the two-stream convolutional neural network have more advantages. Experimental results on UCF101 and HMDB51 data sets show that the proposed method can effectively improve the performance of motion recognition.

A new frog leaping algorithm-oriented fully convolutional neural network for dance motion object saliency detection

Image saliency detection is an important research topic in the field of computer vision. With the traditional saliency detection models, the texture details are not obvious and the edge contour is not complete. The accuracy and recall rate of object detection are low, which are mostly based on the manual features and prior information. With the rise of deep convolutional neural networks, saliency detection has been rapidly developed. However, the existing saliency methods still have some common shortcomings, and it is difficult to uniformly highlight the clear boundary and internal region of the whole object in complex images, mainly because of the lack of sufficient and rich features. In this paper, a new frog leaping algorithm-oriented fully convolutional neural network is proposed for dance motion object saliency detection. The VGG (Visual Geometry Group) model is improved. The final full connection layer is removed, and the jump connection layer is used for the saliency prediction, which can effectively combine the multi-scale information from different convolution layers in the convolutional neural network. Meanwhile, an improved frog leaping algorithm is used to optimize the selection of initial weights during network initialization. In the process of network iteration, the forward propagation loss of convolutional neural network is calculated, and the anomaly weight is corrected by using the improved frog leaping algorithm. When the network satisfies the terminal conditions, the final weight is optimized by one frog leaping to make the network weight further optimization. In addition, the new network can combine high-level semantic information and low-level detail information in a data-driven framework. In order to preserve the unity of the object boundary and inner region effectively, the fully connected conditional random field (CRF) model is used to adjust the obtained saliency feature map. In this paper, the precision recall (PR) curve, F-measure, maximum F-measure, weighted F-measure and mean absolute error (MAE) are tested on six widely used public data sets. Compared with other most advanced and representative methods, the results show that the proposed method achieves better performance and it is superior to most representative methods. The presented method reveals that it has strong robustness for image saliency detection with various scenes, and can make the boundary and inner region of the saliency object more uniform and the detection results more accurate.

Multistage Polymerization Network for Multiperson Pose Estimation

Multiperson pose estimation is an important and complex problem in computer vision. It is regarded as the problem of human skeleton joint detection and solved by the joint heat map regression network in recent years. The key of achieving accurate pose estimation is to learn robust and discriminative feature maps. Although the current methods have made significant progress through interlayer fusion and intralevel fusion of feature maps, few works pay attention to the combination of the two methods. In this paper, we propose a multistage polymerization network (MPN) for multiperson pose estimation. The MPN continuously learns rich underlying spatial information by fusing features within the layers. The MPN also adds hierarchical connections between feature maps at the same resolution for interlayer fusion, so as to reuse low-level spatial information and refine high-level semantic information to obtain accurate keypoint representation. In addition, we observe a lack of connection between the output low-level information and the high-level information. To solve this problem, an effective shuffled attention mechanism (SAM) is proposed. The shuffle aims to promote the cross-channel information exchange between pyramid feature maps, while attention makes a trade-off between the low-level and high-level representations of the output features. As a result, the relationship between the space and the channel of the feature map is further enhanced. Evaluation of the proposed method is carried out on public datasets, and experimental results show that our method has better performance than current methods.

FANet: Feature aggregation network for RGBD saliency detection

The crucial issue in RGBD saliency detection is how to adequately mine and fuse the geometric information and the appearance information contained in depth maps and RGB images, respectively. In this paper, we propose a novel feature aggregation network FANet including a feature extraction module and an aggregation module for RGBD saliency detection. The premier characteristic of FANet is the feature aggregation module consisting of a designed region enhanced module (REM) and a series of deployed hierarchical fusion module (HFM). Specifically, on one hand, the REM provides the powerful capability in differentiating salient objects and background. On the other hand, the HFM is used to gradually integrate high-level semantic information and low-level spatial details, where the K-nearest neighbor graph neural networks (KGNNs) and the non-local module (NLM) are embedded into HFM to dig the geometric information and enhance high-level appearance features, respectively. Extensive experiments on five RGBD datasets show that our model achieves compelling performance against the current 11 state-of-the-art RGBD saliency models.

DAR-Net: Dense Attentional Residual Network for Vehicle Detection in Aerial Images.

With the rapid development of deep learning and the wide usage of Unmanned Aerial Vehicles (UAVs), CNN-based algorithms of vehicle detection in aerial images have been widely studied in the past several years. As a downstream task of the general object detection, there are some differences between the vehicle detection in aerial images and the general object detection in ground view images, e.g., larger image areas, smaller target sizes, and more complex background. In this paper, to improve the performance of this task, a Dense Attentional Residual Network (DAR-Net) is proposed. The proposed network employs a novel dense waterfall residual block (DW res-block) to effectively preserve the spatial information and extract high-level semantic information at the same time. A multiscale receptive field attention (MRFA) module is also designed to select the informative feature from the feature maps and enhance the ability of multiscale perception. Based on the DW res-block and MRFA module, to protect the spatial information, the proposed framework adopts a new backbone that only downsamples the feature map 3 times; i.e., the total downsampling ratio of the proposed backbone is 8. These designs could alleviate the degradation problem, improve the information flow, and strengthen the feature reuse. In addition, deep-projection units are used to reduce the impact of information loss caused by downsampling operations, and the identity mapping is applied to each stage of the proposed backbone to further improve the information flow. The proposed DAR-Net is evaluated on VEDAI, UCAS-AOD, and DOTA datasets. The experimental results demonstrate that the proposed framework outperforms other state-of-the-art algorithms.

FMD-Yolo: An efficient face mask detection method for COVID-19 prevention and control in public

Coronavirus disease 2019 (COVID-19) is a world-wide epidemic and efficient prevention and control of this disease has become the focus of global scientific communities. In this paper, a novel face mask detection framework FMD-Yolo is proposed to monitor whether people wear masks in a right way in public, which is an effective way to block the virus transmission. In particular, the feature extractor employs Im-Res2Net-101 which combines Res2Net module and deep residual network, where utilization of hierarchical convolutional structure, deformable convolution and non-local mechanisms enables thorough information extraction from the input. Afterwards, an enhanced path aggregation network En-PAN is applied for feature fusion, where high-level semantic information and low-level details are sufficiently merged so that the model robustness and generalization ability can be enhanced. Moreover, localization loss is designed and adopted in model training phase, and Matrix NMS method is used in the inference stage to improve the detection efficiency and accuracy. Benchmark evaluation is performed on two public databases with the results compared with other eight state-of-the-art detection algorithms. At IoU = 0.5 level, proposed FMD-Yolo has achieved the best precision AP50 of 92.0% and 88.4% on the two datasets, and AP75 at IoU = 0.75 has improved 5.5% and 3.9% respectively compared with the second one, which demonstrates the superiority of FMD-Yolo in face mask detection with both theoretical values and practical significance.

Deep co-supervision and attention fusion strategy for automatic COVID-19 lung infection segmentation on CT images

Due to the irregular shapes,various sizes and indistinguishable boundaries between the normal and infected tissues, it is still a challenging task to accurately segment the infected lesions of COVID-19 on CT images. In this paper, a novel segmentation scheme is proposed for the infections of COVID-19 by enhancing supervised information and fusing multi-scale feature maps of different levels based on the encoder-decoder architecture. To this end, a deep collaborative supervision (Co-supervision) scheme is proposed to guide the network learning the features of edges and semantics. More specifically, an Edge Supervised Module (ESM) is firstly designed to highlight low-level boundary features by incorporating the edge supervised information into the initial stage of down-sampling. Meanwhile, an Auxiliary Semantic Supervised Module (ASSM) is proposed to strengthen high-level semantic information by integrating mask supervised information into the later stage. Then an Attention Fusion Module (AFM) is developed to fuse multiple scale feature maps of different levels by using an attention mechanism to reduce the semantic gaps between high-level and low-level feature maps. Finally, the effectiveness of the proposed scheme is demonstrated on four various COVID-19 CT datasets. The results show that the proposed three modules are all promising. Based on the baseline (ResUnet), using ESM, ASSM, or AFM alone can respectively increase Dice metric by 1.12%, 1.95%,1.63% in our dataset, while the integration by incorporating three models together can rise 3.97%. Compared with the existing approaches in various datasets, the proposed method can obtain better segmentation performance in some main metrics, and can achieve the best generalization and comprehensive performance.

Research on Surface Defect Detection of Rare-Earth Magnetic Materials Based on Improved SSD

In order to overcome the limitation of manual visual inspection of surface defects of rare-earth magnetic materials and increase production efficiency of traditional rare-earth enterprises, a detection method based on improved SSD (Single Shot Detector) is proposed. The SSD model is improved from two aspects for better performance in the detection of small defects. First of all, the multiscale receptive field module is embedded into the backbone network of the algorithm to improve the feature extraction ability of the model. Secondly, the interlayer feature fusion strategy of bidirectional feature pyramid in PANet (path aggregation network) is integrated into the model. In order to enhance the detection ability of the model, the high-level semantic information is strengthened by an efficient channel attention mechanism. The detection speed of the improved SSD algorithm is 55FPS, and the mAP (mean Average Precision) is up to 83.65%, which is 3.41% higher than of the original SSD algorithm, and the ability to identify small defects is significantly improved.

Object Detection Method Based on Shallow Feature Fusion and Semantic Information Enhancement

In recent years, object detection algorithm based on deep learning has made great progress, but the detection effect is not ideal for small objects detection. Some methods use high-resolution features or enhance shallow features to improve the detection accuracy of small objects. However, using high-resolution features for detection needs higher computational cost, and enhancing shallow features by propagating semantic information from high-level into low-level may bring information aliasing. To address this issue, we propose a novel object detection method based on shallow feature fusion and semantic information enhancement (FFSI). The high-level semantic information is injected into low-level features to guide the enhancement of specific detail information. In order to reduce the information aliasing in shallow features and enhance the receptive field of shallow features, we design two parallel modules: context information enhancement module (CIE) and receptive field enhancement module (RFE). CIE highlights the location of objects by establishing the relationship between local and global context information. RFE enhances the receptive field of shallow features by using dilated convolution to adapt to object detection of different scales, especially small objects. The proposed model is evaluated extensively on PASCAL VOC, and COCO datasets. The experimental results demonstrate that the proposed FFSI model has competitive performance. More importantly, this study reveals that FFSI outperforms the state-of-the-art methods in detecting small objects.