Squeeze-and-Excitation Block Research Articles

Automatic Mask Detection at Public Places

COVID-19's fast spread has resulted in tens of millions of individuals becoming infected around the world. Because there is no specific cure for COVID-19, wearing masks has proven an effective technique of preventing transmission and is mandatory in most public places, resulting in an increase in demand for automatic real-time mask detection systems to replace manual reminders. However, there are just a few investigations on the detection of face masks. The performance of mask detectors must be improved urgently. However, there have been few studies on the detection of face masks. Mask detector performance must be improved immediately. The Properly Wearing Masked Face Detection Dataset (PWMFD) was proposed in this paper, and it includes 9205 photos of mask-wearing samples divided into three groups. Squeeze and Excitation (SE)-YOLOv3, a mask detector with roughly balanced effectiveness and efficiency, was also proposed. by incorporating the SE block into Darknet53, I was able to incorporate the attention mechanism and obtain the relationships between channels, allowing the network to focus more on the relevant feature. To increase the stability of bounding box regression, I used GIoUloss, which can better express the spatial difference between predicted and ground truth boxes. The significant foreground-background class imbalance was solved via focal loss. In addition, I used image augmentation techniques to boost the model's robustness on the challenge. In comparison to YOLOv3, SE-YOLOv3 outperformed YOLOv3 and other state-of-the-art detectors on PWMFD, achieving a higher 8.6 percent mAP while maintaining a comparable detection speed.

Animal Migration Patterns Extraction Based on Atrous-Gated CNN Deep Learning Model

Weather radar data can capture large-scale bird migration information, helping solve a series of migratory ecological problems. However, extracting and identifying bird information from weather radar data remains one of the challenges of radar aeroecology. In recent years, deep learning was applied to the field of radar data processing and proved to be an effective strategy. This paper describes a deep learning method for extracting biological target echoes from weather radar images. This model uses a two-stream CNN (Atrous-Gated CNN) architecture to generate fine-scale predictions by combining the key modules such as squeeze-and-excitation (SE), and atrous spatial pyramid pooling (ASPP). The SE block can enhance the attention on the feature map, while ASPP block can expand the receptive field, helping the network understand the global shape information. The experiments show that in the typical historical data of China next generation weather radar (CINRAD), the precision of the network in identifying biological targets reaches up to 99.6%. Our network can cope with complex weather conditions, realizing long-term and automated monitoring of weather radar data to extract biological target information and provide feasible technical support for bird migration research.

Convolutional squeeze-and-excitation network for ECG arrhythmia detection

Automatic detection of arrhythmia through an electrocardiogram (ECG) is of great significance for the prevention and treatment of cardiovascular diseases. In Convolutional neural network, the ECG signal is converted into multiple feature channels with equal weights through the convolution operation. Multiple feature channels can provide richer and more comprehensive information, but also contain redundant information, which will affect the diagnosis of arrhythmia, so feature channels that contain arrhythmia information should be paid attention to and given larger weight. In this paper, we introduced the Squeeze-and-Excitation (SE) block for the first time for the automatic detection of multiple types of arrhythmias with ECG. Our algorithm combines the residual convolutional module and the SE block to extract features from the original ECG signal. The SE block adaptively enhances the discriminative features and suppresses noise by explicitly modeling the interdependence between the channels, which can adaptively integrate information from different feature channels of ECG. The one-dimensional convolution operation over the time dimension is used to extract temporal information and the shortcut connection of the Se-Residual convolutional module in the proposed model makes the network easier to optimize. Thanks to the powerful feature extraction capabilities of the network, which can effectively extract discriminative arrhythmia features in multiple feature channels, so that no extra data preprocessing including denoising in other methods are need for our framework. It thus improves the working efficiency and keeps the collected biological information without loss. Experiments conducted with the 12-lead ECG dataset of the China Physiological Signal Challenge (CPSC) 2018 and the dataset of PhysioNet/Computing in Cardiology (CinC) Challenge 2017. The experiment results show that our model gains great performance and has great potential in clinical.

Delving deep into spatial pooling for squeeze-and-excitation networks

Squeeze-and-Excitation (SE) blocks have demonstrated significant accuracy gains for state-of-the-art deep architectures by re-weighting channel-wise feature responses. The SE block is an architecture unit that integrates two operations: a squeeze operation that employs global average pooling to aggregate spatial convolutional features into a channel feature, and an excitation operation that learns instance-specific channel weights from the squeezed feature to re-weight each channel. In this paper, we revisit the squeeze operation in SE blocks, and shed lights on why and how to embed rich (both global and local) information into the excitation module at minimal extra costs. In particular, we introduce a simple but effective two-stage spatial pooling process: rich descriptor extraction and information fusion. The rich descriptor extraction step aims to obtain a set of diverse (i.e., global and especially local) deep descriptors that contain more informative cues than global average-pooling. While, absorbing more information delivered by these descriptors via a fusion step can aid the excitation operation to return more accurate re-weight scores in a data-driven manner. We validate the effectiveness of our method by extensive experiments on ImageNet for image classification and on MS-COCO for object detection and instance segmentation. For these experiments, our method achieves consistent improvements over the SENets on all tasks, in some cases, by a large margin.

SAR-U-Net: Squeeze-and-excitation block and atrous spatial pyramid pooling based residual U-Net for automatic liver segmentation in Computed Tomography

Background and objectiveLiver segmentation is an essential prerequisite for liver cancer diagnosis and surgical planning. Traditionally, liver contour is delineated manually by radiologist in a slice-by-slice fashion. However, this process is time-consuming and prone to errors depending on radiologist's experience. In this paper, a modified U-Net based framework is presented, which leverages techniques from Squeeze-and-Excitation (SE) block, Atrous Spatial Pyramid Pooling (ASPP) and residual learning for accurate and robust liver Computed Tomography (CT) segmentation, and the effectiveness of the proposed method was tested on two public datasets LiTS17 and SLiver07. MethodsA new network architecture, called SAR-U-Net was designed, which is grounded in the classical U-Net. Firstly, the SE block is introduced to adaptively extract image features after each convolution in the U-Net encoder, while suppressing irrelevant regions, and highlighting features of specific segmentation task; Secondly, the ASPP is employed to replace the transition layer and the output layer, and acquire multi-scale image information via different receptive fields. Thirdly, to alleviate the gradient vanishment problem, the traditional convolution block is replaced with the residual structures, and thus prompt the network to gain accuracy from considerably increased depth. ResultsIn the LiTS17 database experiment, five popular metrics were used for evaluation, including Dice coefficient, VOE, RVD, ASD and MSD. Compared with other closely related models, the proposed method achieved the highest accuracy. In addition, in the experiment of the SLiver07 dataset, compared with other closely related models, the proposed method achieved the highest segmentation accuracy except for the RVD. ConclusionAn improved U-Net network combining SE, ASPP, and residual structures is developed for automatic liver segmentation from CT images. This new model shows a great improvement on the accuracy compared to other closely related models, and its robustness to challenging problems, including small liver regions, discontinuous liver regions, and fuzzy liver boundaries, is also well demonstrated and validated.

Multispectral LiDAR Point Cloud Classification Using SE-PointNet++

A multispectral light detection and ranging (LiDAR) system, which simultaneously collects spatial geometric data and multi-wavelength intensity information, opens the door to three-dimensional (3-D) point cloud classification and object recognition. Because of the irregular distribution property of point clouds and the massive data volume, point cloud classification directly from multispectral LiDAR data is still challengeable and questionable. In this paper, a point-wise multispectral LiDAR point cloud classification architecture termed as SE-PointNet++ is proposed via integrating a Squeeze-and-Excitation (SE) block with an improved PointNet++ semantic segmentation network. PointNet++ extracts local features from unevenly sampled points and represents local geometrical relationships among the points through multi-scale grouping. The SE block is embedded into PointNet++ to strengthen important channels to increase feature saliency for better point cloud classification. Our SE-PointNet++ architecture has been evaluated on the Titan multispectral LiDAR test datasets and achieved an overall accuracy, a mean Intersection over Union (mIoU), an F1-score, and a Kappa coefficient of 91.16%, 60.15%, 73.14%, and 0.86, respectively. Comparative studies with five established deep learning models confirmed that our proposed SE-PointNet++ achieves promising performance in multispectral LiDAR point cloud classification tasks.

Vehicle classification using a real-time convolutional structure based on DWT pooling layer and SE blocks

Real-time vehicle type classification has become more popular and significant in recent years owing to its potential applications. In this paper, a real-time Convolutional Neural Network is proposed for vehicle type classification. This novel CNN structure is a combination of classic CNN layers along with squeeze-and-excitation (SE) modules and Haar wavelet as the pooling layer. This structure improves the performance of the CNN classifier by emphasizing informative feature maps and decreasing the entropy of the network. A cross-entropy loss function is proposed for better performance. A new pooling method is also introduced based on the Haar transform. Network parameters and the number of layers are selected in such a way to be real-time. Experimental results on two vehicle datasets show that the overall performance of this model, including recognition time and accuracy, is better than the others. For example, using DWT instead of Max-pooling improved the recognition rate on the IRVD dataset from 97.12% to 99.06%. The number of training parameters of the proposed model is about 5 million which is much less than popular networks like VGG (128 m), ResNet152 (58 m), DarkNet (40 m), and Inception-V3 (24 m). This made it very faster so that its recognition time is only 42 ms on CPU which is suitable for real-time applications.

Salient object detection based on adaptive recalibration technique through deep network

Object detection is most required task in computer vision to meet the requirement of autonomous processes. The key challenge in object detection is to separate the foreground known as dominant information of the image from background. In this paper, a new model Squeeze and Excitation Region-Convolutional Neural Network (SER-CNN) is proposed which combines a Squeeze and Excitation (SE) embedded CNN with a Region proposal CNN (R-CNN) to detect salient objects. In SER-CNN, SE blocks are embedded in CNN and this extracts features made up of grand pixels which are accompanied with channel information. R-CNN generates salient map by fusing grand pixels with bounding box techniques to detect objects which is similar to the ground truth. DUT-OMRON, Pascal VOC and SOD dataset applied to find the performance of the SER-CNN. MAE and maxF performance evaluation measures are used for comparing the proposed work with the existing ones.

Real-Time Face Mask Detection Method Based on YOLOv3

The rapid outbreak of COVID-19 has caused serious harm and infected tens of millions of people worldwide. Since there is no specific treatment, wearing masks has become an effective method to prevent the transmission of COVID-19 and is required in most public areas, which has also led to a growing demand for automatic real-time mask detection services to replace manual reminding. However, few studies on face mask detection are being conducted. It is urgent to improve the performance of mask detectors. In this paper, we proposed the Properly Wearing Masked Face Detection Dataset (PWMFD), which included 9205 images of mask wearing samples with three categories. Moreover, we proposed Squeeze and Excitation (SE)-YOLOv3, a mask detector with relatively balanced effectiveness and efficiency. We integrated the attention mechanism by introducing the SE block into Darknet53 to obtain the relationships among channels so that the network can focus more on the important feature. We adopted GIoUloss, which can better describe the spatial difference between predicted and ground truth boxes to improve the stability of bounding box regression. Focal loss was utilized for solving the extreme foreground-background class imbalance. Besides, we performed corresponding image augmentation techniques to further improve the robustness of the model on the specific task. Experimental results showed that SE-YOLOv3 outperformed YOLOv3 and other state-of-the-art detectors on PWMFD and achieved a higher 8.6% mAP compared to YOLOv3 while having a comparable detection speed.

PSigmoid: Improving squeeze-and-excitation block with parametric sigmoid

Squeeze-and-Excitation (SE) Networks won the last ImageNet Large-Scale Visual Recognition Challenge (ILSVRC) classification competition and is very popular in today’s vision community. The SE block is the core of Squeeze-and-Excitation Network (SENet), which adaptively recalibrates channel-wise features and suppresses less useful ones. Since SE blocks can be directly used in existing models and effectively improve performance, SE blocks are widely used in a variety of tasks. In this paper, we propose a novel Parametric Sigmoid (PSigmoid) to enhance the SE block. We named the new module PSigmoid SE (PSE) block. The PSE block can not only suppress features in a channel-wise manner, but also enhance features. We evaluate the performance of our method on four common datasets including CIFAR-10, CIFAR-100, SVHN and Tiny ImageNet. Experimental results show the effectiveness of our method. In addition, we compare the differences between the PSE block and the SE block through a detailed analysis of the configuration. Finally, we use a combination of PSE block and SE block to obtain better performance.

A squeeze and excitation ResNeXt-based deep learning model for Bangla handwritten compound character recognition

In today’s digital era, the recognition of handwritten documents is highly demanding due to the widespread applications. Bangla is one of the most used languages in the world, which consists of 50 basic alphabets. Many compound characters exist in Bangla, which is the combination of two or more basic characters. The recognition of Bangla handwritten characters is a challenging task due to their various sizes, sheer, diversity, lots of turns, similarities between alphabets, and different writing patterns. This paper presents a deep CNN (Convolutional Neural Network) model using the SE-ResNeXt. The squeeze and excitation (SE) blocks are added with existing ResNeXt to address the Bangla handwritten compound character recognition. Usually, CNN extracted the spatial features in the lower-layers and complex features in the upper layers. The SE blocks are added to improve the performance of usual deep CNN by automatically fusing the channel-wise spatial information and inter-channel dependencies through squeezing and excitation within local receptive fields, respectively. To validate the performance of the proposed model, we have utilized Mendeley BanglaLekha-Isolated 2 dataset. The experimental results demonstrate that the proposed model exhibits an average accuracy of 99.82% in recognizing Bangla handwritten compound characters. In addition, the proposed model outperforms the state-of-the-art models by demonstrating higher results.

A sample-level DCNN for music auto-tagging

Deep convolutional neural networks (DCNNs) has been widely used in music auto-tagging which is a multi-label classification task that predicts tags of audio signals. This paper presents a sample-level DCNN for music auto-tagging. The proposed DCNN highlights two components: strided convolutional layer for extracting local feature and reducing temporal dimension, and residual block from WaveNet for preserving input resolution and extracting more complex features. In order to further improve performance, squeeze-and-excitation (SE) block is introduced to the residual block. Under the evaluation metric of Area Under Receiver Operating Characteristic Curve (AUC-ROC) score, experiment results on MagnaTagATune (MTAT) dataset show that the two proposed models achieve 91.47% and 92.76% respectively. Furthermore, our proposed models have slightly surpass the state-of-the-art model SampleCNN with SE block.

An Efficient Memristor-Based Circuit Implementation of Squeeze-and-Excitation Fully Convolutional Neural Networks.

Recently, there has been a surge of interest in applying memristors to hardware implementations of deep neural networks due to various desirable properties of the memristor, such as nonvolativity, multivalue, and nanosize. Most existing neural network circuit designs, however, are based on generic frameworks that are not optimized for memristors. Furthermore, to the best of our knowledge, there are no existing efficient memristor-based implementations of complex neural network operators, such as deconvolutions and squeeze-and-excitation (SE) blocks, which are critical for achieving high accuracy in common medical image analysis applications, such as semantic segmentation. This article proposes convolution-kernel first (CKF), an efficient scheme for designing memristor-based fully convolutional neural networks (FCNs). Compared with existing neural network circuits, CKF enables effective parameter pruning, which significantly reduces circuit power consumption. Furthermore, CKF includes the novel, memristor-optimized implementations of deconvolution layers and SE blocks. Simulation results on real medical image segmentation tasks confirm that CKF obtains up to 56.2% reduction in terms of computations and 33.62-W reduction in terms of power consumption in the circuit after weight pruning while retaining high accuracy on the test set. Moreover, the pruning results can be applied directly to existing circuits without any modification for the corresponding system.

Specific Emitter Identification Based on Software-Defined Radio and Decision Fusion

Specific emitter identification (SEI) uses the unintentional modulation information carried by the emitter waveform, i.e., radio frequency fingerprints, to realize the matching identification of the received signal and its corresponding emitter. We propose an emitter identification scheme based on deep learning (DL). The received signal is subjected to time-varying filtered empirical mode decomposition (tvf-EMD). The obtained intrinsic mode functions (IMFs) are subjected to Hilbert transformation to obtain a 3D-Hilbert spectrum, using amplitude frequency aggregation characteristics obtained by projection to express the nonstationary information of the signal and calculate its bispectral diagonal slice as the second feature. A squeeze-and-excitation block (SEB) is introduced to a convolutional neural network (CNN) to focus on the effective area in the feature map, and support vector machine (SVM) is used to fuse the decision information of the two types of features. We collect four types of steady-state signals on a software-defined radio (SDR) sensor platform based on GNU Radio and universal software radio peripherals (USRP), and study the algorithm performance. The results show that our scheme has strong generalization, high recognition accuracy, and broad application prospects compared with other SEI methods.

CDD-Net: A Context-Driven Detection Network for Multiclass Object Detection

Unlike object detection in natural images that usually achieved great success, remote sensing imagery has its own challenges to detect and localize multiclass objects, such as large-scale change, uncertain direction, and high density. The context information of the objects is very worthwhile for solving these challenges in remote sensing images. In this letter, we propose a context-driven detection network (CDD-Net) to improve the accuracy of multiclass object detection in remote sensing images. For capturing the local neighboring objects and features, a local context feature network (LCFN) is proposed to learn the local context of the region of interest. Meanwhile, a hybrid attention pyramid network (HAPN) is designed, which can steer the focus to more valuable features. The HAPN inserts a squeeze and excitation block (SEB) and three asymmetric convolution blocks (ACBs) in the feature pyramid network (FPN). The experimental results over the DOTA-v1.5 data set demonstrate that the proposed CDD-Net yields promising results.

End-to-end multivariate time series classification via hybrid deep learning architectures

Deep learning has revolutionized many areas, including time series data mining. Multivariate time series classification (MTSC) remained to be a well-known problem in the time series data mining community, due to its availability in various practical applications such as healthcare, finance, geoscience, and bioinformatics. Recently, multivariate long short-term memory with fully convolutional network (MLSTM-FCN) and multivariate attention long short-term memory with fully convolutional network (MALSTM-FCN) have shown superior results over various state-of-the-art methods. So, in this paper, we explore the usage of recurrent neural network (RNN), and its variants, such as bidirectional recurrent neural network (BiRNN), bidirectional long short-term memory (BiLSTM), gated recurrent unit (GRU), and bidirectional gated recurrent unit (BiGRU). We augment these RNN variants separately by replacing long short-term memory (LSTM) in MLSTM-FCN, which is the combination of LSTM, squeeze-and-excitation (SE) block, and fully convolutional network (FCN). Moreover, we integrate the SE block within FCN to leverage its high performance for the MTSC task. The resulting algorithms do not require heavy pre-processing or feature crafting. Thus, they could be easily deployed on real-time systems. We conduct a comprehensive evaluation with a large number of standard datasets and demonstrate that our approaches achieve notable results over the current best MTSC approach.

CT kernel conversions using convolutional neural net for super-resolution with simplified squeeze-and-excitation blocks and progressive learning among smooth and sharp kernels

PurposeComputed tomography (CT) volume sets reconstructed with different kernels are helping to increase diagnostic accuracy. However, several CT volumes reconstructed with different kernels are difficult to sustain, due to limited storage and maintenance issues. A CT kernel conversion method is proposed using convolutional neural networks (CNN). MethodsA total of 3289 CT images from ten patients (five men and five women; mean age, 63.0 ± 8.6 years) were obtained in May 2016 (Somatom Sensation 16, Siemens Medical Systems, Forchheim, Germany). These CT images were reconstructed with various kernels, including B10f (very smooth), B30f (medium smooth), B50f (medium sharp), and B70f (very sharp) kernels. Smooth kernel images were converted into sharp kernel images using super-resolution (SR) network with Squeeze-and-Excitation (SE) blocks and auxiliary losses, and vice versa. In this study, the single-conversion model and multi-conversion model were presented. In case of the single-conversion model, for the one corresponding output image (e.g., B10f to B70), SE-Residual blocks were stacked. For the multi-conversion model, to convert an image into several output images (e.g., B10f to B30f, B50f, and B70f, and vice versa), progressive learning (PL) was employed by calculating auxiliary losses in every four SE-Residual blocks. Through auxiliary losses, the model could learn mutual relationships between different kernel types. The conversion quality was evaluated by the root-mean-square-error (RMSE), structural similarity (SSIM) index and mutual information (MI) between original and converted images. ResultsThe RMSE (SSIM index , MI) of the multi-conversion model was 4.541 ± 0.688 (0.998 ± 0.001 , 2.587 ± 0.137), 27.555 ± 5.876 (0.944 ± 0.021 , 1.735 ± 0.137), 72.327 ± 17.387 (0.815 ± 0.053 , 1.176 ± 0.096), 8.748 ± 1.798 (0.996 ± 0.002 , 2.464 ± 0.121), 9.470 ± 1.772 (0.994 ± 0.003 , 2.336 ± 0.133), and 9.184 ± 1.605 (0.994 ± 0.002 , 2.342 ± 0.138) in conversion between B10f–B30f, B10f–B50f, B10f–B70f, B70f–B50f, B70f–B30f, and B70f–B10f, respectively, which showed significantly better image quality than the conventional model. ConclusionsWe proposed deep learning-based CT kernel conversion using SR network. By introducing simplified SE blocks and PL, the model performance was significantly improved.

Linear Context Transform Block

Squeeze-and-Excitation (SE) block presents a channel attention mechanism for modeling global context via explicitly capturing dependencies across channels. However, we are still far from understanding how the SE block works. In this work, we first revisit the SE block, and then present a detailed empirical study of the relationship between global context and attention distribution, based on which we propose a simple yet effective module, called Linear Context Transform (LCT) block. We divide all channels into different groups and normalize the globally aggregated context features within each channel group, reducing the disturbance from irrelevant channels. Through linear transform of the normalized context features, we model global context for each channel independently. The LCT block is extremely lightweight and easy to be plugged into different backbone models while with negligible parameters and computational burden increase. Extensive experiments show that the LCT block outperforms the SE block in image classification task on the ImageNet and object detection/segmentation on the COCO dataset with different backbone models. Moreover, LCT yields consistent performance gains over existing state-of-the-art detection architectures, e.g., 1.5∼1.7% APbbox and 1.0%∼1.2% APmask improvements on the COCO benchmark, irrespective of different baseline models of varied capacities. We hope our simple yet effective approach will shed some light on future research of attention-based models.

Fully automated segmentation on brain ischemic and white matter hyperintensities lesions using semantic segmentation networks with squeeze-and-excitation blocks in MRI

Abstract Ischemic stroke, a common disease in the elderly, can cause long-term disability and death. Quantitative measurement of brain ischemic lesions at the acute stage is vital for accurate diagnosis and treatment decisions for stroke patients. Currently, the manual segmentation of ischemic lesions is time-consuming, and it is difficult to extract potentially quantifiable information. These limitations can be overcome via deep learning, which has recently become a popular method to segment brain ischemic lesions in MRI. Fully automatic segmentation methods using deep learning quantitatively evaluate the infarct lesions and accurately identify lesions, localize fast lesions, and improve treatment planning. The state-of-the-art methods are variants of U-Net and are fully convolutional networks (FCN); however, these methods have a few limitations, including their inability to capture global features because of the encoder- and decoder-networks. To overcome this limitation, a semantic segmentation method using U-Net with squeeze-and-excitation (SE) blocks is proposed in this study. This method can improve channel-wise information with feature maps compared to a conventional network, to enhance the accuracy of brain lesion segmentation. Patients with acute infarction (N = 429) were retrospectively enrolled in this study with the approval of IRB. Various types of semantic segmentation networks with or without SE blocks were developed, and the accuracies were compared with conventional in-house and commercial software. The Dice similarity coefficient (DSC) of U-Net and Dense U-Net with SE blocks was 85.39 ± 0.84 and 84.23 ± 1.60, respectively. The DSCs increase by 3.5% and 0.9% on average. Accuracies of basic U-Net and Dense U-Net with SE blocks were significantly better than those of conventional image processing methods and those without SE blocks (p-values: 1.0e-08, 2.272e-05, and 0.0003, respectively). To additionally evaluate the methods, a public dataset of white matter hyperintensities (WMHs) segmentation challenge at MICCAI 2017 was used. In the WMH dataset, the DSCs of U-Net with or without SE blocks were 74.64 ± 1.11 and 76.92 ± 0.78, respectively. The DSC increases by 2.3% on average. The semantic segmentation with SE blocks exhibited significantly better performances than those without SE blocks. The solution can be applied to measure various brain segmentation problems, including infarct volume (at the acute stage of stroke patients) and WMH volume.

Robust visual tracking with channel attention and focal loss

Recently, the tracking community leads a fashion of end-to-end feature representation learning for visual tracking. Previous works treat all feature channels and training samples equally during training. This ignores channel interdependencies and foreground–background data imbalance, thus limiting the tracking performance. To tackle these problems, we introduce channel attention and focal loss into the network design to enhance feature representation learning. Specifically, a Squeeze-and-Excitation (SE) block is coupled to each convolutional layer to generate channel attention. Channel attention reflects the channel-wise importance of each feature channel and is used for feature weighting in online tracking. To alleviate the foreground–background data imbalance, we propose a focal logistic loss by adding a modulating factor to the logistic loss, with two tunable focusing parameters. The focal logistic loss down-weights the loss assigned to easy examples in the background area. Both the SE block and focal logistic loss are computationally lightweight and impose only a slight increase in model complexity. Extensive experiments are performed on three challenging tracking datasets including OTB100, UAV123 and TC128. Experimental results demonstrate that the enhanced tracker achieves significant performance improvement while running at a real-time frame-rate of 66 fps.