Foreground-background Imbalance Research Articles

LiViT-Net: A U-Net-like, lightweight Transformer network for retinal vessel segmentation

The intricate task of precisely segmenting retinal vessels from images, which is critical for diagnosing various eye diseases, presents significant challenges for models due to factors such as scale variation, complex anatomical patterns, low contrast, and limitations in training data. Building on these challenges, we offer novel contributions spanning model architecture, loss function design, robustness, and real-time efficacy. To comprehensively address these challenges, a new U-Net-like, lightweight Transformer network for retinal vessel segmentation is presented. By integrating MobileViT+ and a novel local representation in the encoder, our design emphasizes lightweight processing while capturing intricate image structures, enhancing vessel edge precision. A novel joint loss is designed, leveraging the characteristics of weighted cross-entropy and Dice loss to effectively guide the model through the task's challenges, such as foreground-background imbalance and intricate vascular structures. Exhaustive experiments were performed on three prominent retinal image databases. The results underscore the robustness and generalizability of the proposed LiViT-Net, which outperforms other methods in complex scenarios, especially in intricate environments with fine vessels or vessel edges. Importantly, optimized for efficiency, LiViT-Net excels on devices with constrained computational power, as evidenced by its fast performance. To demonstrate the model proposed in this study, a freely accessible and interactive website was established (https://hz-t3.matpool.com:28765?token=aQjYR4hqMI), revealing real-time performance with no login requirements.

Entropy-Based re-sampling method on SAR class imbalance target detection

Detection tasks based on Synthetic aperture radar (SAR) images have been studied widely but severely constrained by the quality of datasets. Meanwhile, both the unperceived category imbalance problem and SAR image discrepancy of multi-class SAR datasets are not fully considered. Researchers usually care about the foreground-background imbalance more than the class imbalance for SAR images. To solve these problems, an entropy-based re-sampling method is proposed in which category imbalance and quality discrepancies of SAR images are both considered. Initially, the relationship is established between variance-weighted image entropy and factors affecting SAR image quality, such as noise, resolution, and density, thereby validating entropy as a robust metric for image quality assessment. Subsequently, quantity penalty scores of categories and difficulty penalty scores of each image are calculated separately to capture the inter-class and intra-class disparities. Next, logarithmic smoothing is employed to avoid overestimation of image difficulty due to the margin effect. Finally, all these scores are combined to generate normalized scores representing the final distribution of the dataset to guide the training process. The proposed approach serves as a plug-and-play strategy for general SAR detection tasks, and experimental results indicate a significant performance improvement. Specifically, detection accuracy in terms of AP for airplanes and bridges (minority classes) in the MSAR dataset is improved by 16.6 % and 13.0 %, respectively, compared to the YOLOv5 baseline, with only a minimal 1.6 % sacrifice in ship detection (majority class). The datasets and the codes can be found at https://www.radars.ac.cn/web/data/getData?dataType=MSAR, https://github.com/Phoenix0qi/Yolov5-entropy-balance/tree/master.

An Anatomy- and Topology-Preserving Framework for Coronary Artery Segmentation.

Coronary artery segmentation is critical for coronary artery disease diagnosis but challenging due to its tortuous course with numerous small branches and inter-subject variations. Most existing studies ignore important anatomical information and vascular topologies, leading to less desirable segmentation performance that usually cannot satisfy clinical demands. To deal with these challenges, in this paper we propose an anatomy- and topology-preserving two-stage framework for coronary artery segmentation. The proposed framework consists of an anatomical dependency encoding (ADE) module and a hierarchical topology learning (HTL) module for coarse-to-fine segmentation, respectively. Specifically, the ADE module segments four heart chambers and aorta, and thus five distance field maps are obtained to encode distance between chamber surfaces and coarsely segmented coronary artery. Meanwhile, ADE also performs coronary artery detection to crop region-of-interest and eliminate foreground-background imbalance. The follow-up HTL module performs fine segmentation by exploiting three hierarchical vascular topologies, i.e., key points, centerlines, and neighbor connectivity using a multi-task learning scheme. In addition, we adopt a bottom-up attention interaction (BAI) module to integrate the feature representations extracted across hierarchical topologies. Extensive experiments on public and in-house datasets show that the proposed framework achieves state-of-the-art performance for coronary artery segmentation.

A Novel Building Extraction Network via Multi-Scale Foreground Modeling and Gated Boundary Refinement

Deep learning-based methods for building extraction from remote sensing images have been widely applied in fields such as land management and urban planning. However, extracting buildings from remote sensing images commonly faces challenges due to specific shooting angles. First, there exists a foreground–background imbalance issue, and the model excessively learns features unrelated to buildings, resulting in performance degradation and propagative interference. Second, buildings have complex boundary information, while conventional network architectures fail to capture fine boundaries. In this paper, we designed a multi-task U-shaped network (BFL-Net) to solve these problems. This network enhances the expression of the foreground and boundary features in the prediction results through foreground learning and boundary refinement, respectively. Specifically, the Foreground Mining Module (FMM) utilizes the relationship between buildings and multi-scale scene spaces to explicitly model, extract, and learn foreground features, which can enhance foreground and related contextual features. The Dense Dilated Convolutional Residual Block (DDCResBlock) and the Dual Gate Boundary Refinement Module (DGBRM) individually process the diverted regular stream and boundary stream. The former can effectively expand the receptive field, and the latter utilizes spatial and channel gates to activate boundary features in low-level feature maps, helping the network refine boundaries. The predictions of the network for the building, foreground, and boundary are respectively supervised by ground truth. The experimental results on the WHU Building Aerial Imagery and Massachusetts Buildings Datasets show that the IoU scores of BFL-Net are 91.37% and 74.50%, respectively, surpassing state-of-the-art models.

FarSeg++: Foreground-Aware Relation Network for Geospatial Object Segmentation in High Spatial Resolution Remote Sensing Imagery.

Geospatial object segmentation, a fundamental Earth vision task, always suffers from scale variation, the larger intra-class variance of background, and foreground-background imbalance in high spatial resolution (HSR) remote sensing imagery. Generic semantic segmentation methods mainly focus on the scale variation in natural scenarios. However, the other two problems are insufficiently considered in large area Earth observation scenarios. In this paper, we propose a foreground-aware relation network (FarSeg++) from the perspectives of relation-based, optimization-based, and objectness-based foreground modeling, alleviating the above two problems. From the perspective of the relations, the foreground-scene relation module improves the discrimination of the foreground features via the foreground-correlated contexts associated with the object-scene relation. From the perspective of optimization, foreground-aware optimization is proposed to focus on foreground examples and hard examples of the background during training to achieve a balanced optimization. Besides, from the perspective of objectness, a foreground-aware decoder is proposed to improve the objectness representation, alleviating the objectness prediction problem that is the main bottleneck revealed by an empirical upper bound analysis. We also introduce a new large-scale high-resolution urban vehicle segmentation dataset to verify the effectiveness of the proposed method and push the development of objectness prediction further forward. The experimental results suggest that FarSeg++ is superior to the state-of-the-art generic semantic segmentation methods and can achieve a better trade-off between speed and accuracy. The code and model are available at: https://github.com/Z-Zheng/FarSeg.

UAVPNet: A balanced and enhanced UAV object detection and pose recognition network

With the increasing popularity of unmanned aerial vehicles (UAVs), accurate positioning and pose recognition of UAVs by target images based on photoelectric detection become a research hotspot. To solve this issue, a multi-scale UAV-Pose dataset consisting of 1400 UAV images is contributed in this paper. In addition, a balanced and enhanced network (UAVPNet) is proposed. UAVPNet has two major features: (1) balanced feature pyramid (BFP) feature fusion structure to improve unbalanced multi-scale features; (2) VarifocalNet detection head to alleviate the foreground-background imbalance. A comparative study demonstrates that UAVPNet is superior to some state-of-the-art object detection models (such as Faster R-CNN-CARAFE, and Yolov8, etc.) in terms of detection accuracy and robustness. Specifically, UAVPNet achieves state-of-the-art 0.885 mAP on the newly-created UAV-Pose dataset, together with nearly 33.49 M parameters, 139.9G FLOPs, and 9.8 FPS. It could fully fulfill the requirements of UAV positioning and pose recognition in the intricate environment.

3D-EddyNet: A Novel Approach for Identifying Three-Dimensional Morphological Features of Mesoscale Eddies in the Ocean

Mesoscale eddies are characterized by swirling currents spanning from tens to hundreds of kilometers in diameter three-dimensional attributes holds paramount significance in driving advancements in both oceanographic research and engineering applications. Nonetheless, a notable absence of models capable of adeptly harnessing the scarcity of high-quality annotated marine data, to efficiently discern the three-dimensional morphological attributes of mesoscale eddies, is evident. To address this limitation, this paper constructs an innovative deep-learning-based model termed 3D-EddyNet, tailored for the precise identification and visualization of mesoscale eddies. In contrast to the prevailing 2D models that remain confined to surface-level data, 3D-EddyNet takes full advantage of three-dimensional convolutions to capture the essential characteristics of eddies. It is specifically tailored for recognizing spatial features within mesoscale eddies, including parameters like position, radius, and depth. The combination of dynamic convolutions and residual networks effectively enhances the model’s performance in a synergistic manner. The model employs the PReLU activation function to tackle gradient vanishing issues and improve convergence rates. It also addresses the challenge of foreground–background imbalance through cross-entropy functions. Additionally, to fine-tune the model’s effectiveness during the training phase, techniques such as random dropblock and batch normalization are skillfully incorporated. Furthermore, we created a training dataset using HYCOM data specifically from the South China Sea region. This dataset allowed for a comprehensive analysis of the spatial-temporal distribution and three-dimensional morphology of the eddies, serving as an assessment of the model’s practical effectiveness. The culmination of this analysis reveals an impressive 20% enhancement over 3D-UNet in detection accuracy, coupled with expedited convergence speed. Notably, the results obtained through our detection using empirical data align closely with those obtained by other scholars. The mesoscale eddies within this specific region unveil a discernible northeast-to-southwest distribution pattern, categorized into three principal morphological classifications: bowl-shaped, olive-shaped, and nearly cylindrical, with the bowl-shaped eddies prominently dominating.

Object Recognition for Millimeter Wave MIMO-SAR Images Based on High-Resolution Feature Recursive Alignment Fusion Network

There are several complex situations in recognizing concealed objects from millimeter wave multiple-input multiple-output synthetic aperture radar (MIMO-SAR) security images, such as incomplete imaging of objects, partially occluded objects, and overlapping objects, which are detrimental to the accurate recognition of concealed objects. To solve these problems, a concealed object detection method based on a high-resolution feature recursive alignment fusion network (HR-FRAFnet) is proposed. The HR-FRAFnet can segment the object area from the grayscale image with complex human background and complete the recognition. The overall architecture of the HR-FRAFnet follows the encoder-decoder framework. Specifically, in the encoder stage, a deep parallel feature extraction network (DPFEN) connects the multi-resolution feature maps in parallel and repeats multi-scale feature fusion. This approach suppresses the background noise flowing and retains more recognizable target characteristics. Then, in the decoder stage, a feature recursive alignment fusion module (FRAFM) is designed to enhance the perception of object edges. The FRAFM effectively improves the segmentation accuracy of objects whereas decreasing the computational complexity of the network. Besides, we employ a combined loss function to alleviate the foreground-background imbalance problem in MIMO-SAR images. Homemade human security screening image datasets are used for evaluation. The experimental results show that the proposed method outperforms existing semantic segmentation methods in Mean Intersection over Union (mIoU) and reduces the incidence of missed and error detection of targets.

Improving building rooftop segmentation accuracy through the optimization of UNet basic elements and image foreground-background balance

Building rooftop segmentation using deep learning techniques is a popular yet challenging area of research in computer vision and remote sensing image processing. While recent studies have developed various deep learning models, there has been less focus on investigating the basic elements of network architecture and foreground-background balance in sampled images. To address this research gap, this study proposes optimizing the UNet basic elements and image foreground-background balance to improve building rooftop segmentation accuracy. The Inria dataset is used for model training and accuracy evaluation. The results show that the impact of network backbone on segmentation accuracy depends on its ability to extract lower-level features and whether it is pretrained by large amounts of data. This study also finds that residual structures of network backbone can negatively affect performance regardless of the network depth. Among the decoding network upsampling strategies, the proposed super-resolution method achieves the best segmentation results, but does not significantly differ from the commonly used bilinear interpolation and deconvolution methods. Additionally, this study proposes a hybrid loss function that considers area, arithmetic, shape, spectral, and texture discrepancy, which further improves the segmentation performance. Finally, this study highlights the significant impact of image foreground-background balance on building rooftop segmentation and proposes an ensemble model to mitigate image foreground-background imbalance and improve segmentation performance.

Fast and Accurate Road Crack Detection Based on Adaptive Cost-Sensitive Loss Function.

Numerous detection problems in computer vision, including road crack detection, suffer from exceedingly foreground-background imbalance. Fortunately, modification of loss function appears to solve this puzzle once and for all. In this article, we propose a pixel-based adaptive weighted cross-entropy (WCE) loss in conjunction with Jaccard distance to facilitate high-quality pixel-level road crack detection. Our work profoundly demonstrates the influence of loss functions on detection outcomes and sheds light on the sophisticated consecutive improvements in the realm of crack detection. Specifically, to verify the effectiveness of the proposed loss, we conduct extensive experiments on four public databases, that is, CrackForest, AigleRN, Crack360, and BJN260. Compared to the vanilla WCE, the proposed loss significantly speeds up the training process while retaining the performance.

SPANet: Spatial Adaptive Convolution Based Content-Aware Network for Aerial Image Semantic Segmentation

Semantic segmentation of remote sensing images encounters four significant difficulties: 1) complex backgrounds, 2) large-scale differences, 3) numerous small objects, and 4) extreme foreground-background imbalance. However, the existing generic semantic segmentation models mainly focus on the modeling context information and rarely focus on these four issues. This article presents an enhanced remote sensing image semantic segmentation framework to solve these problems through the Hierarchical Atrous Pyramid module (HASP) and spatial-adaptive convolution based FPN decoder framework (SPA-FPN). On the one hand, HASP solved the problem of complex backgrounds and large-scale differences by further enlarging the receptive field of the network through the cascade of atrous convolution with various rates. On the other hand, spatial adaptive convolution is embedded in FPN decoder framework step by step to solve the problems of numerous small objects and extreme foreground-background imbalance. Besides, a boundary-based loss function is constructed to help the network optimize the relevant segmentation results. Extensive experiments over iSAID and ISPRS Vaihingen datasets reflect the superiority of the presented structure to conventional state-of-the-art semantic segmentation approaches. The code will be accessible.

Context Information Refinement for Few-Shot Object Detection in Remote Sensing Images

Recently, few-shot object detection based on fine-tuning has attracted much attention in the field of computer vision. However, due to the scarcity of samples in novel categories, obtaining positive anchors for novel categories is difficult, which implicitly introduces the foreground–background imbalance problem. It is difficult to identify foreground objects from complex backgrounds due to various object sizes and cluttered backgrounds. In this article, we propose a novel context information refinement few-shot detector (CIR-FSD) for remote sensing images. In particular, we design a context information refinement (CIR) module to extract discriminant context features. This module uses dilated convolutions and dense connections to capture rich context information from different receptive fields and then uses a binary map as the supervision label to refine the context information. In addition, we improve the region proposal network (RPN). Concretely, the RPN is fine-tuned on novel categories, and the constraint of non-maximum suppression (NMS) is relaxed, which can obtain more positive anchors for novel categories. Experiments on two remote sensing public datasets show the effectiveness of our detector.

Residual objectness for imbalance reduction

As most object detectors rely on dense candidate samples to cover objects, they have always suffered from the extreme imbalance between very few foreground samples and numerous background samples during training, i.e., the foreground-background imbalance. Although several resampling and reweighting schemes (e.g., OHEM, Focal Loss, GHM) have been proposed to alleviate the imbalance, they are usually heuristic with multiple hyper-parameters, which is difficult to generalize on different object detectors and datasets. In this paper, we propose a novel Residual Objectness (ResObj) mechanism that adaptively learns how to address the foreground-background imbalance problem in object detection. Specifically, we first formulate the imbalance problems on all object classes as an imbalance problem on an “objectness” class. Then, we design multiple cascaded objectness estimators with residual connections for that objectness class to progressively distinguish the foreground samples from background samples. With our residual objectness mechanism, object detectors can learn how to address the foreground-background problem in an end-to-end way, rather than rely on hand-crafted resampling or reweighting schemes. Extensive experiments on the COCO benchmark demonstrate the effectiveness and compatibility of our method for various object detectors: the RetinaNet-ResObj, YOLOv3-ResObj and FasterRCNN-ResObj achieve relative 3%∼4% Average Precision (AP) improvements compared with their vanilla models, respectively.

An Anchor-Free Method Based on Adaptive Feature Encoding and Gaussian-Guided Sampling Optimization for Ship Detection in SAR Imagery

Recently, deep-learning methods have yielded rapid progress for object detection in synthetic aperture radar (SAR) imagery. It is still a great challenge to detect ships in SAR imagery due to ships’ small size and confusable detail feature. This article proposes a novel anchor-free detection method composed of two modules to deal with these problems. First, for the lack of detailed information on small ships, we suggest an adaptive feature-encoding module (AFE), which gradually fuses deep semantic features into shallow layers and realizes the adaptive learning of the spatial fusion weights. Thus, it can effectively enhance the external semantics and improve the representation ability of small targets. Next, for the foreground–background imbalance, the Gaussian-guided detection head (GDH) is introduced according to the idea of soft sampling and exploits Gaussian prior to assigning different weights to the detected bounding boxes at different locations in the training optimization. Moreover, the proposed Gauss-ness can down-weight the predicted scores of bounding boxes far from the object center. Finally, the effect of the detector composed of the two modules is verified on the two SAR ship datasets. The results demonstrate that our method can effectively improve the detection performance of small ships in datasets.

Improving Semantic Segmentation in Aerial Imagery via Graph Reasoning and Disentangled Learning

Semantic segmentation in aerial imagery is still an important, yet challenging task due to the complex characteristics of remote-sensing data. The critical issues consist of: 1) extreme foreground–background imbalance; 2) large intra-class variance; and 3) arbitrary-oriented, dense, and small objects. The above challenges make it unlikely to model the effective global interdependencies of semantic heterogeneous regions. Besides, general semantic segmentation methods suffer from feature ambiguity due to the joint feature learning paradigm, leading to inferior detail information. In this article, we propose an improved semantic segmentation framework to tackle these problems via graph reasoning (GR) and disentangled learning. On the one hand, a simple, yet effective GR unit is introduced to implement coordinate-interaction space mapping and perform relation reasoning over the graph. It can be deployed on the feature pyramid network (FPN) to exploit cross-stage multi-scale information. On the other hand, we propose a so- called disentangled learning paradigm to explicitly model the foreground and boundary objects, instantiated as foreground prior estimation (FPE) and boundary alignment (BA). The indication of the intermediate feature can be effectively emphasized to enhance the discriminative abilities of the network. Extensive experiments over iSAID, ISPRS Vaihingen, and the general Cityscapes datasets demonstrate the effectiveness and efficiency of the proposed framework over other state-of-the-art semantic segmentation methods.

Robust Vehicle Detection in High-Resolution Aerial Images With Imbalanced Data

Vehicle detection in images from unmanned aerial vehicles (UAVs) plays an important role in traffic surveillance and urban planning due to the popularity of UAVs. However, the class imbalance problem is an important factor that restricts the performance of vehicle detectors. There are two types of class imbalance in UAV images, i.e., foreground-background imbalance and foreground–foreground imbalance. For anchor-based single stage detector, as many ground truths cannot be assigned to corresponding anchors because of low intersection over union, it makes the foreground-background imbalance problem more severe. Therefore, we propose a novel bag-based single-stage detector, which treats each position on the feature map as a bag. A simple and adaptive definition of bags is proposed along with the positive sample definition method, which is utilized to ensure more ground truths can be assigned to proper bags. In addition, we utilize online hard example mining method to control the proportion of positive and negative samples during the training process. To address the foreground–foreground imbalance, we propose a novel data augmentation algorithm, which allows us to create appropriate visual context for under-represented class. Extensive experiments demonstrate the superiority of the proposed algorithm, compared with other state-of-the-art solutions. <p xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Impact Statement—Recently, unmanned aerial vehicles (UAVs) are widely used in intelligent transportation due to their low price and high flexibility, which makes vehicle detection in UAV images important for automatically gathering of traffic information. However, the class imbalance problem, which is common in object detection where some classes have far fewer frequencies in the dataset, has an adverse effect on the performance of vehicle detectors. The data augmentation method and deep learning based vehicle detector proposed in this article are able to reduce the negative impact and improve detection performance by at least 1.27% in mean average precision index. In addition, compared with algorithms with similar detection performance, our method is at least 15 ms faster. The proposed method can benefit users in a wide variety of applications including UAV transportation, traffic surveillance, and urban planning.

Introspective False Negative Prediction for Black-Box Object Detectors in Autonomous Driving.

Object detection plays a critical role in autonomous driving, but current state-of-the-art object detectors will inevitably fail in many driving scenes, which is unacceptable for safety-critical automated vehicles. Given the complexity of the real traffic scenarios, it is impractical to guarantee zero detection failure; thus, online failure prediction is of crucial importance to mitigate the risk of traffic accidents. Of all the failure cases, False Negative (FN) objects are most likely to cause catastrophic consequences, but little attention has been paid to the online FN prediction. In this paper, we propose a general introspection framework that can make online prediction of FN objects for black-box object detectors. In contrast to existing methods which rely on empirical assumptions or handcrafted features, we facilitate the FN feature extraction by an introspective FN predictor we designed in this framework. For this purpose, we extend the original concept of introspection to object-wise FN predictions, and propose a multi-branch cooperation mechanism to address the distinct foreground-background imbalance problem of FN objects. The effectiveness of the proposed framework is verified through extensive experiments and analysis, and the results show that our method successfully predicts the FN objects with 81.95% precision for 88.10% recall on the challenging KITTI Benchmark, and effectively improves object detection performance by taking FN predictions into consideration.

Is Heuristic Sampling Necessary in Training Deep Object Detectors?

To train accurate deep object detectors under the extreme foreground-background imbalance, heuristic sampling methods are always necessary, which either re-sample a subset of all training samples (hard sampling methods, e.g. biased sampling, OHEM), or use all training samples but re-weight them discriminatively (soft sampling methods, e.g. Focal Loss, GHM). In this paper, we challenge the necessity of such hard/soft sampling methods for training accurate deep object detectors. While previous studies have shown that training detectors without heuristic sampling methods would significantly degrade accuracy, we reveal that this degradation comes from an unreasonable classification gradient magnitude caused by the imbalance, rather than a lack of re-sampling/re-weighting. Motivated by our discovery, we propose a simple yet effective Sampling-Free mechanism to achieve a reasonable classification gradient magnitude by initialization and loss scaling. Unlike heuristic sampling methods with multiple hyperparameters, our Sampling-Free mechanism is fully data diagnostic, without laborious hyperparameters searching. We verify the effectiveness of our method in training anchor-based and anchor-free object detectors, where our method always achieves higher detection accuracy than heuristic sampling methods on COCO and PASCAL VOC datasets. Our Sampling-Free mechanism provides a new perspective to address the foreground-background imbalance. Our code is released at https://github.com/ChenJoya/sampling-free.

Limited sliding network: Fine‐grained target detection on electrical infrastructure for power transmission line surveillance

SummaryBecause of its small size, low local contrast, and much interference, the field image of fine‐grained equipment taken from power transmission line surveillance is hard to be sustained by the traditional small target detection technique, which requires the manual extraction of features, making it difficult to accurately detect micro‐fine‐grained equipment. The deep learning‐based algorithms have prospective application but require abundant data to guarantee performance and tackle the problem of foreground–background imbalance. This paper develops an effective pipeline, i.e., limited sliding network (LSNet), to detect the small and fine‐grained defects on equipment in power transmission line infrastructure. The model firstly performs the regional analysis on the entire image to determine the potential target locations. The feature extraction and classification on the potential location image blocks are further performed by the VGG‐style model for the dense target locations, and the nonmaximum suppression method is finally applied to locate the target. On the other hand, a specific training method is also developed to better deal with a wide range imbalances of positive and negative samples. The proposed method achieves the detection mean average precision (mAP) rate of 98.66% on the real datasets, while limiting the computational overhead of hardware.

Automatic and Robust Object Detection in X-Ray Baggage Inspection Using Deep Convolutional Neural Networks

For the purpose of ensuring public security, automatic inspection of X-ray scanners has been deployed at the entry points of many public places to detect dangerous objects. However, current surveillance systems cannot function without human supervision and intervention. In this article, we propose an effective method using deep convolutional neural networks to detect objects during X-ray baggage inspection. As a first step, a large amount of training data is generated by a specific data augmentation technique. Second, a feature enhancement module is used to improve feature extraction capabilities. Then, in order to address the foreground–background imbalance in the region proposal network, focal loss is adopted. Third, the multiscale fused region of interest is utilized to obtain more robust proposals. Finally, soft nonmaximum suppression is adopted to alleviate overlaps in baggage detection. As compared with existing algorithms, the proposed method proves that it is more accurate and robust when dealing with densely cluttered backgrounds during X-ray baggage inspection.