Channel Attention Module Research Articles

Improved YOLOv5-Based Real-Time Road Pavement Damage Detection in Road Infrastructure Management

Deep learning has enabled a straightforward, convenient method of road pavement infrastructure management that facilitates a secure, cost-effective, and efficient transportation network. Manual road pavement inspection is time-consuming and dangerous, making timely road repair difficult. This research showcases You Only Look Once version 5 (YOLOv5), the most commonly employed object detection model trained on the latest benchmark Road Damage Dataset, Road Damage Detection 2022 (RDD 2022). The RDD 2022 dataset includes four common types of road pavement damage, namely vertical cracks, horizontal cracks, alligator cracks, and potholes. This paper presents an improved deep neural network model based on YOLOv5 for real-time road pavement damage detection in photographic representations of outdoor road surfaces, making it an indispensable tool for efficient, real-time, and cost-effective road infrastructure management. The YOLOv5 model has been modified to incorporate several techniques that improve its accuracy and generalization performance. These techniques include the Efficient Channel Attention module (ECA-Net), label smoothing, the K-means++ algorithm, Focal Loss, and an additional prediction layer. In addition, a 1.9% improvement in mean average precision (mAP) and a 1.29% increase in F1-Score were attained by the model in comparison to YOLOv5s, with an increment of 1.1 million parameters. Moreover, a 0.11% improvement in mAP and 0.05% improvement in F1 score was achieved by the proposed model compared to YOLOv8s while having 3 million fewer parameters and 12 gigabytes fewer Giga Floating Point Operation per Second (GFlops).

PKDN: Prior Knowledge Distillation Network for bronchoscopy diagnosis

Bronchoscopy plays a crucial role in diagnosing and treating lung diseases. The deep learning-based diagnostic system for bronchoscopic images can assist physicians in accurately and efficiently diagnosing lung diseases, enabling patients to undergo timely pathological examinations and receive appropriate treatment. However, the existing diagnostic methods overlook the utilization of prior knowledge of medical images, and the limited feature extraction capability hinders precise focus on lesion regions, consequently affecting the overall diagnostic effectiveness. To address these challenges, this paper proposes a prior knowledge distillation network (PKDN) for identifying lung diseases through bronchoscopic images. The proposed method extracts color and edge features from lesion images using the prior knowledge guidance module, and subsequently enhances spatial and channel features by employing the dynamic spatial attention module and gated channel attention module, respectively. Finally, the extracted features undergo refinement and self-regulation through feature distillation. Furthermore, decoupled distillation is implemented to balance the importance of target and non-target class distillation, thereby enhancing the diagnostic performance of the network. The effectiveness of the proposed method is validated on the bronchoscopic dataset provided by Harbin Medical University Cancer Hospital, which consists of 2,029 bronchoscopic images from 200 patients. Experimental results demonstrate that the proposed method achieves an accuracy of 94.78% and an AUC of 98.17%, outperforming other methods significantly in diagnostic performance. These results indicate that the computer-aided diagnostic system based on PKDN provides satisfactory accuracy in diagnosing lung diseases during bronchoscopy.

Deep learning model for displacement monitoring of super high arch dams based on measured temperature data

Dam displacement, an important indicator for the health monitoring of dam structures, can effectively reflect its operational status. Displacement prediction models based on measured data are currently an important tool for dam safety monitoring. However, a significant portion of current models rely on statistical or shallow machine-based approaches, posing challenges in extracting comprehensive and impactful features from environmental factors. To address the above problems, this paper proposes a novel deep learning model that combines Inception architectures with residual connections (Inception-ResNet) and Gate Recurrent Units (GRU). This model employs improved Inception-ResNet blocks with channel attention and spatial attention modules to extract features from dam deformation-related environmental factors sequences at multiple scales. Subsequently, GRU is utilized to learn from long-term dependencies. The proposed model fully combines the remarkable feature extraction capability of the Inception-ResNet block with the learning capability of GRU for long-term dependencies. The availability of the proposed model is tested with measured data of a super high arch dam. The experimental results show that the proposed model outperforms two typical shallow machine learning methods and two typical deep learning models in the four typical monitoring points selected, which demonstrates convincingly that the proposed model is able to predict dam deformation with high accuracy and robustness for dam structure safety monitoring.

Camellia oleifera Fruit Detection Algorithm in Natural Environment Based on Lightweight Convolutional Neural Network

At present, Camellia oleifera fruit harvesting relies on manual labor with low efficiency, while mechanized harvesting could result in bud damage because flowering and fruiting are synchronized. As a prerequisite, rapid detection and identification are urgently needed for high accuracy and efficiency with simple models to realize selective and intelligent harvesting. In this paper, a lightweight detection algorithm YOLOv5s-Camellia based on YOLOv5s is proposed. First, the network unit of the lightweight network ShuffleNetv2 was used to reconstruct the backbone network, and thereby the number of computations and parameters of the model was reduced to increase the running speed for saving computational costs. Second, to mitigate the impact of the lightweight improvement on model detection accuracy, three efficient channel attention (ECA) modules were introduced into the backbone network to enhance the network’s attention to fruit features, and the Concat operation in the neck network was replaced by the Add operation with fewer parameters, which could increase the amount of information under features while maintaining the same number of channels. Third, the Gaussian Error Linear Units (GELU) activation function was introduced to improve the nonlinear characterization ability of the network. In addition, to improve the ability of the network to locate objects in the natural environment, the penalty index was redefined to optimize the bounding box loss function, which can improve the convergence speed and regression accuracy. Furthermore, the final experimental results showed that this model possesses 98.8% accuracy, 5.5 G FLOPs computation, and 6.3 MB size, and the detection speed reached 60.98 frame/s. Compared with the original algorithm, the calculation amount, size, and parameters were reduced by 65.18%, 56.55%, and 57.59%, respectively. The results can provide a technical reference for the development of a Camellia oleifera fruit-harvesting robot.

Attention-Mechanism-Based Models for Unconstrained Face Recognition with Mask Occlusion

Masks cover most areas of the face, resulting in a serious loss of facial identity information; thus, how to alleviate or eliminate the negative impact of occlusion is a significant problem in the field of unconstrained face recognition. Inspired by the successful application of attention mechanisms and capsule networks in computer vision, we propose ECA-Inception-Resnet-Caps, which is a novel framework based on Inception-Resnet-v1 for learning discriminative face features in unconstrained mask-wearing conditions. Firstly, Squeeze-and-Excitation (SE) modules and Efficient Channel Attention (ECA) modules are applied to Inception-Resnet-v1 to increase the attention on unoccluded face areas, which is used to eliminate the negative impact of occlusion during feature extraction. Secondly, the effects of the two attention mechanisms on the different modules in Inception-Resnet-v1 are compared and analyzed, which is the foundation for further constructing the ECA-Inception-Resnet-Caps framework. Finally, ECA-Inception-Resnet-Caps is obtained by improving Inception-Resnet-v1 with capsule modules, which is explored to increase the interpretability and generalization of the model after reducing the negative impact of occlusion. The experimental results demonstrate that both attention mechanisms and the capsule network can effectively enhance the performance of Inception-Resnet-v1 for face recognition in occlusion tasks, with the ECA-Inception-Resnet-Caps model being the most effective, achieving an accuracy of 94.32%, which is 1.42% better than the baseline model.

An attention-based multi-scale convolution network for intelligent underwater acoustic signal recognition

Underwater acoustic signal identification is always an extremely challenging task due to the intricate marine environments and background noise interference. Existing techniques typically focus on handling only one type of specific features for recognition, overlooking a comprehensive representation of distinct vessel attributes and struggling to handle intricate scenarios. To address these challenges, we employ a band-pass filter bank for feature extraction. This process extracts multiple representations of the same data, each concentrating on distinct frequency components. Subsequently, the filtered multi-view representation is channeled into an attention-based multi-scale convolution network (AMSC-Net) for recognition purposes. AMSC-Net integrates a sequence of attention-based multi-scale convolution (AMSC) blocks to capture pivotal features tied to identifying diverse ship-radiated audio categories. Specifically, a set of parallel residual convolution layers with different kernels are designed to capture multi-scale features, and an adaptive channel attention module is followed to highlight dominant parts within global features. To enrich training data in terms of quantity and diversity, a data augmentation strategy is implemented during the training phase. Our proposals are evaluated on two real-world underwater acoustic signal datasets. Experimental results firmly establish the superiority of our proposed AMSC-Net, which substantially outperforms competitive DNN-based methods. Moreover, our proposals exhibit robustness across different noise levels, achieving an overall accuracy of 95.87% and 95.39% for the two datasets at 0 dB signal-to-noise ratio (SNR), respectively.

CAPN: a Combine Attention Partial Network for glove detection.

Accidents caused by operators failing to wear safety gloves are a frequent problem at electric power operation sites, and the inefficiency of manual supervision and the lack of effective supervision methods result in frequent electricity safety accidents. To address the issue of low accuracy in glove detection with small-scale glove datasets. This article proposes a real-time glove detection algorithm using video surveillance to address these issues. The approach employs transfer learning and an attention mechanism to enhance detection average precision. The key ideas of our algorithm are as follows: (1) introducing the Combine Attention Partial Network (CAPN) based on convolutional neural networks, which can accurately recognize whether gloves are being worn, (2) combining channel attention and spatial attention modules to improve CAPN's ability to extract deeper feature information and recognition accuracy, and (3) using transfer learning to transfer human hand features in different states to gloves to enhance the small sample dataset of gloves. Experimental results show that the proposed network structure achieves high performance in terms of detection average precision. The average precision of glove detection reached 96.59%, demonstrating the efficacy of CAPN.

AFCANet: An adaptive feature concatenate attention network for multi-focus image fusion

For multi-focus image fusion, the existing deep learning based methods cannot effectively learn the texture features and semantic information of the source image to generate high-quality fused images. Thus, we develop a new adaptive feature concatenate attention network named AFCANet, which adaptively learns cross-layer features and retains the texture features and semantic information of images to generate visually appealing fully focused images. In AFCANet, the encoder-decoder network is used as the backbone network. In the unsupervised training stage, an adaptive cross-layer skip connection mode is designed, and a cross-layer adaptive coordinate attention module is built to acquire meaningful information from the image along with ignoring unimportant information to obtain a better image fusion effect. In addition, in the middle of the encoder-decoder network, we also introduce an effective channel attention module to fully learn the output of the encoder, and accelerate network convergence. In the inference stage, we apply the pixel-based spatial frequency fusion rules to fuse the adaptive features learned by the encoder, which can successfully combine the texture and semantic information of the image and produce a more precise decision map. Extensive experiments on public datasets and the HBU-CVMDSP dataset show that our AFCANet can effectively improve the accuracy of the decision map in the focus and defocus regions, as well as improve the ability to retain the abundant details and edge features of the source image.

CNN and Transformer Fusion for Remote Sensing Image Semantic Segmentation

Semantic segmentation of remote sensing images has been widely used in environmental protection, geological disaster discovery, and natural resource assessment. With the rapid development of deep learning, convolutional neural networks (CNNs) have dominated semantic segmentation, relying on their powerful local information extraction capabilities. Due to the locality of convolution operation, it can be challenging to obtain global context information directly. However, Transformer has excellent potential in global information modeling. This paper proposes a new hybrid convolutional and Transformer semantic segmentation model called CTFuse, which uses a multi-scale convolutional attention module in the convolutional part. CTFuse is a serial structure composed of a CNN and a Transformer. It first uses convolution to extract small-size target information and then uses Transformer to embed large-size ground target information. Subsequently, we propose a spatial and channel attention module in convolution to enhance the representation ability for global information and local features. In addition, we also propose a spatial and channel attention module in Transformer to improve the ability to capture detailed information. Finally, compared to other models used in the experiments, our CTFuse achieves state-of-the-art results on the International Society of Photogrammetry and Remote Sensing (ISPRS) Vaihingen and ISPRS Potsdam datasets.

Partial channel pooling attention beats convolutional attention

The widely used convolutional attention blocks (e.g., Squeeze-and-Excitation attention, Coordinate attention) all use an indirect convolution layer for modeling between channels, lacking direct interaction. In this article, we present a lightweight channel attention mechanism, partial channel pooling attention (PPA), for handling object detection and segmentation tasks. While minimizing the redundancy of channels and pixels, complementary information is achieved through direct interaction between channels. Specifically, we pool the feature maps into pooling modules (pooling blocks) of size KxK. Each pixel after pooling is considered a feature of each channel. In this way, our attention mechanism (PPA) is lightweight and can be directly embedded into existing networks. Our channel attention module selectively emphasizes interdependent channel mappings through the direct interaction of information between channels rather than indirectly through convolutional layers. Through extensive experiments in object detection and semantic segmentation, our PPA performs better on various types of base models. For instance, adding PPA achieves training throughput comparable to adding Coordinate attention but has twice the improvement effect of adding Coordinate attention in COCO detection and semantic segmentation.

MSA-YOLOv5: Multi-scale attention-based YOLOv5 for automatic detection of acute ischemic stroke from multi-modality MRI images

Background and objectiveAcute ischemic stroke (AIS) is a common neurological disorder characterized by the sudden onset of cerebral ischemia, leading to functional impairments. Swift and precise detection of AIS lesions is crucial for stroke diagnosis and treatment but poses a significant challenge. This study aims to leverage multimodal fusion technology to combine complementary information from various modalities, thereby enhancing the detection performance of AIS target detection models. MethodsIn this retrospective study of AIS, we collected data from 316 AIS patients and created a multi-modality magnetic resonance imaging (MRI) dataset. We propose a Multi-Scale Attention-based YOLOv5 (MSA-YOLOv5), targeting challenges such as small lesion size and blurred borders at low resolutions. Specifically, we augment YOLOv5 with a prediction head to detect objects at various scales. Next, we replace the original prediction head with a Multi-Scale Swin Transformer Prediction Head (MS-STPH), which reduces computational complexity to linear levels and enhances the ability to detect small lesions. We incorporate a Second-Order channel attention (SOCA) module to adaptively rescale channel features by employing second-order feature statistics for more discriminative representations. Finally, we further validate the effectiveness of our method using the ISLES 2022 dataset. ResultsOn our in-house AIS dataset, MSA-YOLOv5 achieves a 79.0% mAP0.5, substantially surpassing other single-stage models. Compared to two-stage models, it maintains a comparable performance level while significantly reducing the number of parameters and resolution. On the ISLES 2022 dataset, MSA-YOLOv5 attains an 80.0% mAP0.5, outperforming other network models by a considerable margin. MS-STPH and SOCA modules can significantly increase mAP0.5 by 2.7% and 1.9%, respectively. Visualization interpretability results show that the proposed MSA-YOLOv5 restricts high attention in the small regions of AIS lesions. ConclusionsThe proposed MSA-YOLOv5 is capable of automatically and effectively detecting acute ischemic stroke lesions in multimodal images, particularly for small lesions and artifacts. Our enhanced model reduces the number of parameters while improving detection accuracy. This model can potentially assist radiologists in providing more accurate diagnosis, and enable clinicians to develop better treatment plans.

Multiscale Convolutional Attention-based Residual Network Expression Recognition

<p>Expression recognition has wide application in the fields of distance education and clinical medicine. In response to the problems of insufficient feature extraction ability of expression recognition models in current research, and the deeper the depth of the model, the more serious the loss of useful information, a residual network model with multi-scale convolutional attention is proposed. This model mainly takes the residual network as the main body, adds normalization layer and channel attention mechanism, so as to extract useful image information at multiple scales, and incorporates the Inception module and channel attention module into the residual network to enhance the feature extraction ability of the model and to prevent the loss of more useful information due to too deep network, and to improve the generalization performance of the model. From results of lots of experiments we can see that the recognition accuracy of the model in FER+ and CK+ datasets reaches 87.80% and 99.32% respectively, with better recognition performance and robustness.</p> <p> </p>

Multiple Myeloma Prognosis From PET Images: Deep Survival Losses and Contrastive Pretraining

Objective: Diagnosis and follow-up of multiple myeloma (MM) patients involve analysing full-body Positron Emission Tomography (PET) images. Towards assisting the analysis, there has been an increased interest in machine learning methods linking PET radiomics with survival analysis. Despite deep-learning’s success in other fields, its adaptation to survival faces several challenges. Our goal is to design a deep-learning approach to predict the progression-free survival of MM patients from PET lesion images. Methods: We study three aspects of such deep-learning approach: i) the loss function: we review existing and propose new losses for survival analysis based on contrastive triplet learning, ii) Pre-training: We conceive two pre-training strategies to cope with the relatively small datasets, based on patch classification and triplet loss embedding. iii) the architecture: we study the contribution of spatial and channel attention modules. Results: Our approach is validated on data from two prospective clinical studies, improving the c-index over baseline methods, notably thanks to the channel attention module and the introduced pre-training methods. Conclusion and significance: We propose for the first time an end-to-end deep-learning approach, M2P2, to predict the progression-free survival of MM patients from PET lesion images. We introduce two contrastive learning approaches, never used before for survival analysis.

Unsupervised Pixel-Level Detection of Rail Surface Defects Using Multistep Domain Adaptation

Detection of rail surface defects using deep learning largely relies on supervision with pixel-level ground truths, where the detection accuracy suffers due to unseen image domains. In order to streamline the laborious labeling process and accurately identify defects on rail surfaces a U-Net-based network that integrates spatial and channel attention (SA) modules, multistep domain adaptation (MSDA), and progressive histogram matching (PHM) is developed in the context of pixel-level defects detection in this research. MSDA with an unlabeled intermediate dataset is used to reduce the domain gap underlying different domains. To further improve the adaptation of the transferred model, the PHM method and the self-training method are employed to enhance detection accuracy. Trained in the source domain with ground truths, the deep learning model is then tested in the target domain with no annotations. Experimental results on the RSDD dataset and the Rail-Joint dataset demonstrate the applicability and effectiveness of the proposed approach. By alleviating the domain discrepancy, the developed model drastically enhances the average precision (AP) score of unsupervised detection from 0.103 to 0.861, which shows competitive performance compared with the AP score of 0.895 by the state-of-the-art supervised models. The novelty of this research lies in the development of a hybrid deep learning approach that is able to achieve competitive detection accuracy with a low overlap ratio between domains and few data labeling.

C-SwinIR: Dental CT super-resolution reconstruction fused with SwinIR

Super-resolution reconstruction (SR) of dental computed tomography (Dental CT) images is a innovative and challenging task. To address the limitations of Dental CT in obtaining high-resolution (HR) images due to equipment constraints and noise interference, we propose a Dental CT SR method called C-SwinIR based on SwinIR. Firstly, the self-calibrated convolutions network (SCNet) is introduced to solve the problem of detail loss in shallow feature graphs and improve the ability to recover details. Subsequently, the cross-shaped windows (CSWin) self-attention transformer structure is used to replace the original transformer structure, which improves the ability of the model to obtain context information. Eventually, the integration of efficient channel attention (ECA-Net) module effectively realizes the local cross-channel interaction, accelerates the model convergence and solves the problem of gradient explosion in training. Experimental results show that our proposed method is superior to the original SwinIR network by achieving a higher peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) by 0.287 and 0.003 respectively. Additionally, it decreases mean squared error (MSE) by 1.108. By using this method, clinicians can obtain clearer details and textures from Dental CT, which effectively assist in diagnosis.

GSM-Net: A global sequence modelling network for the segmentation of short axis CINE MRI images.

Atrial Fibrillation (AF) is a disease where the atria fail to properly contract but quiver instead, due to the abnormal electrical activity of the atrial tissue. In AF patients, anatomical and functional parameters of the left atrium (LA) largely differ from that of healthy people due to LA remodelling, which can continue in many cases after the catheter ablation treatment. Therefore, it is important to follow up with AF patients to detect any recurrence. LA segmentation masks obtained from short-axis CINE MRI images are used as the gold standard for the quantification of LA parameters. Thick slices of CINE MRI images hinder the use of 3D networks for segmentation while 2D architectures often fail to model inter-slice dependencies. This study presents GSM-Net which approximates 3D networks with effective modelling of inter-slice similarities with two new modules: global slice sequence encoder (GSSE) and sequence dependent channel attention module (SdCAt). In contrast to previous work modelling only local inter-slice similarities, GSSE also models global spatial dependencies across slices. SdCAt generates a distribution of attention weights over MRI slices per channel, to better trace characteristic changes in the size of the LA or other structures across slices. We found that GSM-Net outperforms previous methods on LA segmentation and helps to identify AF recurrence patients. We believe that GSM-Net can be used as an automatic tool to estimate LA parameters such as ejection fraction to identify AF, and to follow up with patients after treatment to detect any recurrence.

An Efficient and Lightweight Model for Automatic Modulation Classification: A Hybrid Feature Extraction Network Combined with Attention Mechanism

This paper proposes a hybrid feature extraction convolutional neural network combined with a channel attention mechanism (HFECNET-CA) for automatic modulation recognition (AMR). Firstly, we designed a hybrid feature extraction backbone network. Three different forms of convolution kernels were used to extract features from the original I/Q sequence on three branches, respectively, learn the spatiotemporal features of the original signal from different “perspectives” through the convolution kernels with different shapes, and perform channel fusion on the output feature maps of the three branches to obtain a multi-domain mixed feature map. Then, the deep features of the signal are extracted by connecting multiple convolution layers in the time domain. Secondly, a plug-and-play channel attention module is constructed, which can be embedded into any feature extraction layer to give higher weight to the more valuable channels in the output feature map to achieve the purpose of feature correction for the output feature map. The experimental results on the RadiomL2016.10A dataset show that the designed HFECNET-CA has higher recognition accuracy and fewer trainable parameters compared to other networks. Under 20 SNRs, the average recognition accuracy reached 63.92%, and the highest recognition accuracy reached 93.64%.

High-Resolution Remote Sensing Image Segmentation Algorithm Based on Improved Feature Extraction and Hybrid Attention Mechanism

Segmentation of high-resolution remote sensing images is one of the hottest topics in deep learning. Compared to ordinary images, high-resolution remote sensing images possess characteristics such as higher intra-class diversity and lower inter-class separability. Additionally, the objects in these images are complex and have smaller sizes. Aiming at the classical segmentation network in remote sensing images, there are some problems, such as inaccurate edge object segmentation, inconsistent segmentation of different types of objects, low detection accuracy, and a high false detection rate. This paper proposes a new hybrid attention model (S-CA), a new coordinate efficient channel attention module (C-ECA), and a new small-target feature extraction network (S-FE). The S-CA model enhances important spatial and channel features in shallow layers, allowing for more detailed feature extraction. The C-ECA model utilizes convolutional layers to capture complex dependencies between variations, thereby better capturing feature information at each position and reducing redundancy in feature channels. The S-FE network can capture the local feature information of different targets more effectively. It enhances the recognition and classification capabilities of various targets and improves the detection rate of small targets. The algorithm is used for segmentation in high-resolution remote sensing images. Experiments were conducted on the public dataset GID-15 based on Gaofen-2 satellite remote sensing images. The experimental results demonstrate that the improved DeepLabV3+ segmentation algorithm for remote sensing images achieved a mean intersection over union (mIoU), mean pixel accuracy (mPA), and mean precision (mP) of 91.6%, 96.1%, and 95.5%, respectively. The improved algorithm is more effective than current mainstream segmentation networks.

Global-and-local aware network for low-light image enhancement

Photos taken under nighttime or backlit conditions often suffer from complex and unpredictable degradation, such as low visibility, messy noise, and distorted color. Previous methods mainly focused on global brightness and contrast while ignoring structural and textural details, or they handled the fusion of features without adequately considering their intrinsic association, resulting in incomplete feature representations. To address this issue, we propose a global-and-local aware network (GLAN) by projecting the features into the frequency domain and incorporating them in a knowledge-sharing manner. This method effectively integrates the global modeling capability of the transformer and the local sensitivity of the convolutional neural network to represent structure and texture. First, the global branch, which is comprised of transformer blocks, performs feature extraction under the global receptive field, while the local branch constructs multi-scale features to provide local fine-grained details. Then, we design a novel adaptive multi-scale feature block (AMSFB) that deploys channel split operation to decrease the calculation amount. To better learn the channel and spatial correlations of intermediate features, we introduce a multi-scale channel attention module (MSCAM) and a pixel attention module (PAM) into the AMSFB. Finally, a frequency-aware interaction module (FAIM) is developed for bidirectional information supplementation, which builds feature descriptors simultaneously covering low-frequency and high-frequency information based on the discrete cosine transform (DCT). Through extensive quantitative and qualitative experiments, our method can achieve competitive results compared with over ten state-of-the-art image enhancement methods on eight benchmark datasets.

Radar Timing Range–Doppler Spectral Target Detection Based on Attention ConvLSTM in Traffic Scenes

With the development of autonomous driving and the emergence of various intelligent traffic scenarios, object detection technology based on deep learning is more and more widely applied to real traffic scenarios. Commonly used detection devices include LiDAR and cameras. Since the implementation of traffic scene target detection technology requires mass production, the advantages of millimeter-wave radar have emerged, such as low cost and no interference from the external environment. The performance of LiDAR and cameras is greatly reduced due to their sensitivity to light, which affects target detection at night and in bad weather. However, millimeter-wave radar can overcome the influence of these harsh environments and has a great auxiliary effect on safe driving on the road. In this work, we propose a deep-learning-based object detection method considering the radar range–Doppler spectrum in traffic scenarios. The algorithm uses YOLOv8 as the basic architecture, makes full use of the time series characteristics of range–Doppler spectrum data in traffic scenarios, introduces the ConvLSTM network, and exerts the ability to process time series data. In order to improve the model’s ability to detect small objects, an efficient and lightweight Efficient Channel Attention (ECA) module is introduced. Through extensive experiments, our model shows better performance on two publicly available radar datasets, CARRADA and RADDet, compared to other state-of-the-art methods. Compared with other mainstream methods that can only achieve 30–60% mAP performance when the IOU is 0.3, our model can achieve 74.51% and 75.62% on the RADDet and CARRADA datasets, respectively, and has better robustness and generalization ability.