Fusion Block Research Articles

Density‐Aware Diffusion Model for Efficient Image Dehazing

AbstractExisting image dehazing methods have made remarkable progress. However, they generally perform poorly on images with dense haze, and often suffer from unsatisfactory results with detail degradation or color distortion. In this paper, we propose a density‐aware diffusion model (DADM) for image dehazing. Guided by the haze density, our DADM can handle images with dense haze and complex environments. Specifically, we introduce a density‐aware dehazing network (DADNet) in the reverse diffusion process, which can help DADM gradually recover a clear haze‐free image from a haze image. To improve the performance of the network, we design a cross‐feature density extraction module (CDEModule) to extract the haze density for the image and a density‐guided feature fusion block (DFFBlock) to learn the effective contextual features. Furthermore, we introduce an indirect sampling strategy in the test sampling process, which not only suppresses the accumulation of errors but also ensures the stability of the results. Extensive experiments on popular benchmarks validate the superior performance of the proposed method. The code is released in https://github.com/benchacha/DADM.

Noise-assisted hybrid attention networks for low-dose PET and CT denoising.

Positron emission tomography (PET) and computed tomography (CT) play a vital role in tumor-related medical diagnosis, assessment, and treatment planning. However, full-dose PET and CT pose the risk of excessive radiation exposure to patients, whereas low-dose images compromise image quality, impacting subsequent tumor recognition and diseasediagnosis. To solve such problems, we propose a Noise-Assisted Hybrid Attention Network (NAHANet) to reconstruct full-dose PET and CT images from low-dose PET (LDPET) and CT (LDCT) images to reduce patient radiation risks while ensuring the performance of subsequent tumor recognition. NAHANet contains two branches: the noise feature prediction branch (NFPB) and the cascaded reconstruction branch. Among them, NFPB providing noise features for the cascade reconstruction branch. The cascaded reconstruction branch comprises a shallow feature extraction module and a reconstruction module which contains a series of cascaded noise feature fusion blocks (NFFBs). Among these, the NFFB fuses the features extracted from low-dose images with the noise features obtained by NFPB to improve the feature extraction capability. To validate the effectiveness of the NAHANet method, we performed experiments using two public available datasets: the Ultra-low Dose PET Imaging Challenge dataset and Low Dose CT Grand Challengedataset. As a result, the proposed NAHANet achieved higher performance on common indicators. For example, on the CT dataset, the PSNR and SSIM indicators were improved by 4.1dB and 0.06 respectively, and the rMSE indicator was reduced by 5.46 compared with the LDCT; on the PET dataset, the PSNR and SSIM was improved by 3.37dB and 0.02, and the rMSE was reduced by 9.04 compared with theLDPET. This paper proposes a transformer-based denoising algorithm, which utilizes hybrid attention to extract high-level features of low dose images and fuses noise features to optimize the denoising performance of the network, achieving good performance improvements on low-dose CT and PETdatasets.

SFDiff: Diffusion model with sufficient spatial‐Fourier frequency information interaction for low‐light image enhancement

AbstractDiffusion models are increasingly applied in low‐light image enhancement tasks due to their exceptional capability to model data distributions, but most current methods focus only on the original pixel space and neglect the potential of Fourier frequency information. In this article, SFDiff is proposed, a novel low‐light image enhancement method that integrates Fourier frequency information into the diffusion process. Specifically, Fourier transforms are applied at both the image and feature levels to separately enhance the amplitude and phase components, which restores global illumination degradation and positional information. Then a Spatial‐Frequency Fusion (SFF) block is used to fully integrate and interact with the information across spatial and frequency domains. Since illumination degradation is primarily manifested in the amplitude component, a loss function based on maximum likelihood learning is employed to constrain the amplitude component at each step of the sampling process, ensuring that the reverse process maintains an optimal trajectory. Owing to the streamlined network design and the fact that the Fourier transform requires no extra parameters, SFDiff achieves a reduction in parameters of over compared to several state‐of‐the‐art (SOTA) diffusion models and delivers high‐quality enhancement results on multiple real‐world datasets. The code is available at https://github.com/MrWan001/SFDiff.

Microexpression Recognition Method Based on ADP-DSTN Feature Fusion and Convolutional Block Attention Module

Microexpressions are subtle facial movements that occur within an extremely brief time frame, often revealing suppressed emotions. These expressions hold significant importance across various fields, including security monitoring and human–computer interaction. However, the accuracy of microexpression recognition is severely constrained by the inherent characteristics of these expressions. To address the issue of low detection accuracy regarding the subtle features present in microexpressions’ facial action units, this paper proposes a microexpression action unit detection algorithm, Attention-embedded Dual Path and Shallow Three-stream Networks (ADP-DSTN), that incorporates an attention-embedded dual path and a shallow three-stream network. First, an attention mechanism was embedded after each Bottleneck layer in the foundational Dual Path Networks to extract static features representing subtle texture variations that have significant weights in the action units. Subsequently, a shallow three-stream 3D convolutional neural network was employed to extract optical flow features that were particularly sensitive to temporal and discriminative characteristics specific to microexpression action units. Finally, the acquired static facial feature vectors and optical flow feature vectors were concatenated to form a fused feature vector that encompassed more effective information for recognition. Each facial action unit was then trained individually to address the issue of weak correlations among the facial action units, thereby facilitating the classification of microexpression emotions. The experimental results demonstrated that the proposed method achieved great performance across several microexpression datasets. The unweighted average recall (UAR) values were 80.71%, 89.55%, 44.64%, 80.59%, and 88.32% for the SAMM, CASME II, CAS(ME)3, SMIC, and MEGC2019 datasets, respectively. The unweighted F1 scores (UF1) were 79.32%, 88.30%, 43.03%, 81.12%, and 88.95%, respectively. Furthermore, when compared to the benchmark model, our proposed model achieved better performance with lower computational complexity, characterized by a Floating Point Operations (FLOPs) value of 1087.350 M and a total of 6.356 × 106 model parameters.

Degradation-Guided Multi-Modal Fusion Network for Depth Map Super-Resolution

Depth map super-resolution (DSR) is a technique aimed at restoring high-resolution (HR) depth maps from low-resolution (LR) depth maps. In this process, color images are commonly used as guidance to enhance the restoration procedure. However, the intricate degradation of LR depth poses a challenge, and previous image-guided DSR approaches, which implicitly model the degradation in the spatial domain, often fall short of producing satisfactory results. To address this challenge, we propose a novel approach called the Degradation-Guided Multi-modal Fusion Network (DMFNet). DMFNet explicitly characterizes the degradation and incorporates multi-modal fusion in both spatial and frequency domains to improve the depth quality. Specifically, we first introduce the deep degradation regularization loss function, which enables the model to learn the explicit degradation from the LR depth maps. Simultaneously, DMFNet converts the color images and depth maps into spectrum representations to provide comprehensive multi-domain guidance. Consequently, we present the multi-modal fusion block to restore the depth maps by leveraging both the RGB-D spectrum representations and the depth degradation. Extensive experiments demonstrate that DMFNet achieves state-of-the-art (SoTA) performance on four benchmarks, namely the NYU-v2, Middlebury, Lu, and RGB-D-D datasets.

An Improved YOLOv8-Based Foreign Detection Algorithm for Transmission Lines

This research aims to overcome three major challenges in foreign object detection on power transmission lines: data scarcity, background noise, and high computational costs. In the improved YOLOv8 algorithm, the newly introduced lightweight GSCDown (Ghost Shuffle Channel Downsampling) module effectively captures subtle image features by combining 1 × 1 convolution and GSConv technology, thereby enhancing detection accuracy. CSPBlock (Cross-Stage Partial Block) fusion enhances the model’s accuracy and stability by strengthening feature expression and spatial perception while maintaining the algorithm’s lightweight nature and effectively mitigating the issue of vanishing gradients, making it suitable for efficient foreign object detection in complex power line environments. Additionally, PAM (pooling attention mechanism) effectively distinguishes between background and target without adding extra parameters, maintaining high accuracy even in the presence of background noise. Furthermore, AIGC (AI-generated content) technology is leveraged to produce high-quality images for training data augmentation, and lossless feature distillation ensures higher detection accuracy and reduces false positives. In conclusion, the improved architecture reduces the parameter count by 18% while improving the mAP@0.5 metric by a margin of 5.5 points when compared to YOLOv8n. Compared to state-of-the-art real-time object detection frameworks, our research demonstrates significant advantages in both model accuracy and parameter size.

KAN-HyperMP: An Enhanced Fault Diagnosis Model for Rolling Bearings in Noisy Environments

Rolling bearings often produce non-stationary signals that are easily obscured by noise, particularly in high-noise environments, making fault detection a challenging task. To address this challenge, a novel fault diagnosis approach based on the Kolmogorov–Arnold Network-based Hypergraph Message Passing (KAN-HyperMP) model is proposed. The KAN-HyperMP model is composed of three key components: a neighbor feature aggregation block, a feature fusion block, and a KANLinear block. Firstly, the neighbor feature aggregation block leverages hypergraph theory to integrate information from more distant neighbors, aiding in the reduction of noise impact, even when nearby neighbors are severely affected. Subsequently, the feature fusion block combines the features of these higher-order neighbors with the target node’s own features, enabling the model to capture the complete structure of the hypergraph. Finally, the smoothness properties of B-spline functions within the Kolmogorov–Arnold Network (KAN) are employed to extract critical diagnostic features from noisy signals. The proposed model is trained and evaluated on the Southeast University (SEU) and Jiangnan University (JNU) Datasets, achieving accuracy rates of 99.70% and 99.10%, respectively, demonstrating its effectiveness in fault diagnosis under both noise-free and noisy conditions.

DWTN: deep wavelet transform network for lightweight single image deraining

On rainy days the uncertainty of the shape and distribution of rain streaks can cause the images captured by RGB image-based measurement tools to be blurred and distorted. Thanks to the wavelet transform ability to provide spatial and frequency domain information about an image and its multidirectional and multiscale nature, it is widely used in traditional image enhancement methods. In image deraining the distribution of rain streaks is not only related to spatial domain features but is also closely related to frequency domain spatial features. However, deep learning-based rain removal models mainly rely on the spatial features of the image, and RGB data can hardly distinguish rain marks from image details, which leads to the loss of crucial image information during rain removal. We have developed a lightweight single-image rain removal model called the deep wavelet transform network (DWTN) to address this limitation. This method separates image details from rain images and can more effectively remove rain marks. The proposed DWTN has three significant contributions. First, DWTN uses the feature components after the wavelet transform as the input to the model and assigns a separate frequency-aware enhancement block (FAEB) to each element. These blocks focus on specific frequency features that benefit the rain removal task. Second, we introduce a frequency feature fusion block (FFFB) that fuses different wavelet components to reduce noise and enhance the image background through a channel attention mechanism while attenuating rain streaks. Finally, we design a spatial feature enhancement block (SFEB), which uses a spatial attention mechanism to calibrate the spatial position of features to improve the rain removal performance. We evaluate the performance of DWTN using PSNR and SSIM on four synthetic datasets and NIQE and BRISQUE on two real datasets. The results of the evaluation of the six datasets and at least four performance metrics show that the proposed DWTN is superior to existing methods.

DELR-Net: a network for 3D multimodal medical image registration in more lightweight application scenarios.

3D multimodal medical image deformable registration plays a significant role in medical image analysis and diagnosis. However, due to the substantial differences between images of different modalities, registration is challenging and requires high computational costs. Deep learning-based registration methods face these challenges. The primary aim of this paper is to design a 3D multimodal registration network that ensures high-quality registration results while reducing the number of parameters. This study designed a Dual-Encoder More Lightweight Registration Network (DELR-Net). DELR-Net is a low-complexity network that integrates Mamba and ConvNet. The State Space Sequence Module and the Dynamic Large Kernel block are used as the main components of the dual encoders, while the Dynamic Feature Fusion block is used as the main component of the decoder. This study conducted experiments on 3D brain MR images and abdominal MR and CT images. Compared to existing registration methods, DELR-Net achieved better registration results while maintaining a lower number of parameters. Additionally, generalization experiments on other modalities showed that DELR-Net has superior generalization capabilities. DELR-Net significantly improves the limitations of 3D multimodal medical image deformable registration, achieving better registration performance with fewer parameters.

Adaptive Multi-Feature Attention Network for Image Dehazing

Currently, deep-learning-based image dehazing methods occupy a dominant position in image dehazing applications. Although many complicated dehazing models have achieved competitive dehazing performance, effective methods for extracting useful features are still under-researched. Thus, an adaptive multi-feature attention network (AMFAN) consisting of the point-weighted attention (PWA) mechanism and the multi-layer feature fusion (AMLFF) is presented in this paper. We start by enhancing pixel-level attention for each feature map. Specifically, we design a PWA block, which aggregates global and local information of the feature map. We also employ PWA to make the model adaptively focus on significant channels/regions. Then, we design a feature fusion block (FFB), which can accomplish feature-level fusion by exploiting a PWA block. The FFB and PWA constitute our AMLFF. We design an AMLFF, which can integrate three different levels of feature maps to effectively balance the weights of the inputs to the encoder and decoder. We also utilize the contrastive loss function to train the dehazing network so that the recovered image is far from the negative sample and close to the positive sample. Experimental results on both synthetic and real-world images demonstrate that this dehazing approach surpasses numerous other advanced techniques, both visually and quantitatively, showcasing its superiority in image dehazing.

A paralleled CNN and Transformer network for PPG-based cuff-less blood pressure estimation

Cuff-less blood pressure measurement plays a pivotal part in hypertension diagnosis and prevention. By employing deep learning algorithms, blood pressure estimation applying photoplethysmography (PPG) signals has achieved remarkable improvements and attracted growing attention. However, most existing methods utilize the PPG signal emphasized on localized feature learning, often neglecting the global features in the temporal data. To tackle the issue, we propose a novel architecture. It leverages a parallel processing approach by using Convolutional Neural Network for extracting local features and simultaneously employing Transformer to capture global features. The feature fusion block integrates the features using spatial and channel attention. The blood pressure values are regressed by two fully connected layers. To verify the performance of our method, we conduct evaluations on the Medical Information Mart for Intensive Care (MIMIC) database. On the MIMIC database, our algorithm demonstrates exceptional performance. For Diastolic Pressure, Mean Arterial Pressure, and Systolic Pressure, the mean absolute errors are 2.36 mmHg, 2.03 mmHg, and 4.44 mmHg, respectively. This performance aligns with the rigorous criteria required by an Association for the Advancement of Medical Instrumentation (AAMI) and receives a Grade A rating in accordance with the British Hypertension Society (BHS) standard.

AFC-Unet: Attention-fused full-scale CNN-transformer unet for medical image segmentation

In the field of medical image segmentation, although U-Net has achieved significant achievements, it still exposes some inherent disadvantages when dealing with complex anatomical structures and small targets, such as inaccurate target localization, blurry edges, and insufficient integration of contextual information. To address these challenges, this study proposes the Attention-Fused Full-Scale CNN-Transformer Unet (AFC-Unet), aiming to effectively overcome the limitations of traditional U-Net through multi-scale feature fusion, attention mechanisms, and CNN-Transformer hybrid modules. Specifically, we adopt an encoder–decoder architecture, incorporating full-scale feature block fusion and pyramid sampling modules to enhance the model’s ability to recognize fine to overall structural features by integrating cross-level multi-scale features. We propose the Multi-feature Fusion Attention Gates (MFAG) module, which focuses on and highlights discriminative information of potential lesions and key anatomical boundaries, effectively suppressing irrelevant background interference. We design a module Convolutional Hybrid Attention Transformer (CHAT) that integrates CNN and Transformer to address the shortcomings of traditional single models in handling long-range dependencies and global context understanding. Experimental results on three datasets of different scales demonstrate that the model’s segmentation performance for medical images surpasses state-of-the-art models, showcasing high generalization ability.

MEEAFusion: Multi-Scale Edge Enhancement and Joint Attention Mechanism Based Infrared and Visible Image Fusion.

Infrared and visible image fusion can integrate rich edge details and salient infrared targets, resulting in high-quality images suitable for advanced tasks. However, most available algorithms struggle to fully extract detailed features and overlook the interaction of complementary features across different modal images during the feature fusion process. To address this gap, this study presents a novel fusion method based on multi-scale edge enhancement and a joint attention mechanism (MEEAFusion). Initially, convolution kernels of varying scales were utilized to obtain shallow features with multiple receptive fields unique to the source image. Subsequently, a multi-scale gradient residual block (MGRB) was developed to capture the high-level semantic information and low-level edge texture information of the image, enhancing the representation of fine-grained features. Then, the complementary feature between infrared and visible images was defined, and a cross-transfer attention fusion block (CAFB) was devised with joint spatial attention and channel attention to refine the critical supplemental information. This allowed the network to obtain fused features that were rich in both common and complementary information, thus realizing feature interaction and pre-fusion. Lastly, the features were reconstructed to obtain the fused image. Extensive experiments on three benchmark datasets demonstrated that the MEEAFusion proposed in this research has considerable strengths in terms of rich texture details, significant infrared targets, and distinct edge contours, and it achieves superior fusion performance.

A Multi-Branch Feature Extraction Residual Network for Lightweight Image Super-Resolution

Single-image super-resolution (SISR) seeks to elucidate the mapping relationships between low-resolution and high-resolution images. However, high-performance network models often entail a significant number of parameters and computations, presenting limitations in practical applications. Therefore, prioritizing a light weight and efficiency becomes crucial when applying image super-resolution (SR) to real-world scenarios. We propose a straightforward and efficient method, the Multi-Branch Feature Extraction Residual Network (MFERN), to tackle lightweight image SR through the fusion of multi-information self-calibration and multi-attention information. Specifically, we have devised a Multi-Branch Residual Feature Fusion Module (MRFFM) that leverages a multi-branch residual structure to succinctly and effectively fuse multiple pieces of information. Within the MRFFM, we have designed the Multi-Scale Attention Feature Fusion Block (MAFFB) to adeptly extract features via convolution and self-calibration attention operations. Furthermore, we introduce a Dual Feature Calibration Block (DFCB) to dynamically fuse feature information using dynamic weight values derived from the upper and lower branches. Additionally, to overcome the limitation of convolution in solely extracting local information, we incorporate a Transformer module to effectively integrate global information. The experimental results demonstrate that our MFERN exhibits outstanding performance in terms of model parameters and overall performance.

Image Inpainting via Correlated Multi-Resolution Feature Projection.

With the advancement in image editing applications, image inpainting is gaining more attention due to its ability to recover corrupted images efficiently. Also, the existing methods for image inpainting either use two-stage coarse-to-fine architectures or single-stage architectures with a deeper network. On the other hand, shallow network architectures lack the quality of results and the methods with remarkable inpainting quality have high complexity in terms of number of parameters or average run time. Despite the improvement in the inpainting quality, these methods still lack the correlated local and global information. In this work, we propose a single-stage multi-resolution generator architecture for image inpainting with moderate complexity and superior outcomes. Here, a multi-kernel non-local (MKNL) attention block is proposed to merge the feature maps from all the resolutions. Further, a feature projection block is proposed to project features of MKNL to respective decoder for effective reconstruction of image. Also, a valid feature fusion block is proposed to merge encoder skip connection features at valid region and respective decoder features at hole region. This ensures that there will not be any redundant feature merging while reconstruction of image. Effectiveness of the proposed architecture is verified on CelebA-HQ Liu, et al. 2015, Karras et al. 2017, and Places2 Zhou et al. 2018 datasets corrupted with publicly available NVIDIA mask dataset Liu et al. 2018. The detailed ablation study, extensive result analysis, and application of object removal prove the robustness of the proposed method over existing state-of-the-art methods for image inpainting.

Fusing spatial and frequency features for compositional zero-shot image classification

Compositional Zero-Shot Learning (CZSL) is a particular Zero-Shot Learning (ZSL) task that aims to utilize known concepts (e.g., states and objects) to identify novel state-object compositions for Image Classification. Previous works have primarily focused on disentangling concept compositions or exploring the complex interactions between the states and objects while neglecting the critical fact that the inference of many states and compositions is related to different frequency components, which should be analyzed from the global perspective. Therefore, we propose a Spatial-frequency Feature Fusion Network (SFFNet) to introduce a new branch that utilizes a frequency-domain filtering encoder to enhance key frequency components and capture non-local interactions adaptively. Besides, we also find that the widely used backbone in conventional CZSL settings behaves superior in perceiving local features. Thus, we construct a fusion block to combine both strengths to capture the local and non-local information. In addition, the traditional one-hot ground-truth distribution in the training phase does not reflect the accurate relationships between compositions, so we propose a composition-relation based label distribution regularization to encourage the model to actively learn the inner relationships between compositions, and extend this method to construct unseen composition pseudo distribution to further enhance the model’s generalization ability to unseen compositions. Extensive experiments and detailed analysis are conducted on three popular datasets, and the results show that our method can achieve state-of-the-art performance, which reveals its superiority in identifying novel compositions. Code is available at https://github.com/lisuyi/SFFNet_czsl.

Multi-View Feature Fusion and Rich Information Refinement Network for Semantic Segmentation of Remote Sensing Images

Semantic segmentation is currently a hot topic in remote sensing image processing. There are extensive applications in land planning and surveying. Many current studies combine Convolutional Neural Networks (CNNs), which extract local information, with Transformers, which capture global information, to obtain richer information. However, the fused feature information is not sufficiently enriched and it often lacks detailed refinement. To address this issue, we propose a novel method called the Multi-View Feature Fusion and Rich Information Refinement Network (MFRNet). Our model is equipped with the Multi-View Feature Fusion Block (MAFF) to merge various types of information, including local, non-local, channel, and positional information. Within MAFF, we introduce two innovative methods. The Sliding Heterogeneous Multi-Head Attention (SHMA) extracts local, non-local, and positional information using a sliding window, while the Multi-Scale Hierarchical Compressed Channel Attention (MSCA) leverages bar-shaped pooling kernels and stepwise compression to obtain reliable channel information. Additionally, we introduce the Efficient Feature Refinement Module (EFRM), which enhances segmentation accuracy by interacting the results of the Long-Range Information Perception Branch and the Local Semantic Information Perception Branch. We evaluate our model on the ISPRS Vaihingen and Potsdam datasets. We conducted extensive comparison experiments with state-of-the-art models and verified that MFRNet outperforms other models.

BCMFIFuse: A Bilateral Cross-Modal Feature Interaction-Based Network for Infrared and Visible Image Fusion

The main purpose of infrared and visible image fusion is to produce a fusion image that incorporates less redundant information while incorporating more complementary information, thereby facilitating subsequent high-level visual tasks. However, obtaining complementary information from different modalities of images is a challenge. Existing fusion methods often consider only relevance and neglect the complementarity of different modalities’ features, leading to the loss of some cross-modal complementary information. To enhance complementary information, it is believed that more comprehensive cross-modal interactions should be provided. Therefore, a fusion network for infrared and visible fusion is proposed, which is based on bilateral cross-feature interaction, termed BCMFIFuse. To obtain features in images of different modalities, we devise a two-stream network. During the feature extraction, a cross-modal feature correction block (CMFC) is introduced, which calibrates the current modality features by leveraging feature correlations from different modalities in both spatial and channel dimensions. Then, a feature fusion block (FFB) is employed to effectively integrate cross-modal information. The FFB aims to explore and integrate the most discriminative features from the infrared and visible image, enabling long-range contextual interactions to enhance global cross-modal features. In addition, to extract more comprehensive multi-scale features, we develop a hybrid pyramid dilated convolution block (HPDCB). Comprehensive experiments on different datasets reveal that our method performs excellently in qualitative, quantitative, and object detection evaluations.

Data-driven predictive maintenance framework considering the multi-source information fusion and uncertainty in remaining useful life prediction

The existing equipment maintenance methods mainly separate the two related phases of prediction and predictive maintenance (PdM) by looking at remaining useful life (RUL) prediction without considering maintenance or optimizing maintenance schedules based on the given prediction information. To address this issue, we propose a framework based on a multi-stream attention fusion network with quantile regression model and deep reinforcement learning (DRL). In this novel framework, the multi-stream attention fusion block is used to comprehensively capture the operating status of industrial equipment and eliminate the extracted duplicate information. Quantile regression is employed to obtain the probabilistic RUL prediction results with uncertainty expression. We further formulate the PdM problem for DRL, where maintenance actions are triggered based on the estimates of the RUL distribution. To illustrate the superiority of our framework, we compare it with some state-of-the-art models using a public data and our private data. The experimental results indicate that our method exhibits high accuracy and stability in bearing RUL prediction and predictive replacement.

A deep learning–based surrogate model for spatial-temporal temperature field prediction in subway tunnel fires via CFD simulation

The smoke temperature and toxic gas generated by subway tunnel fires threaten trapped personnel and hinder rescue work. In this study, with a view to ensuring that smoke movement is acquired quickly and contributing to the future development of subway tunnel fires to guide emergency rescue, a deep learning surrogate model based on a transposed convolution neural network (TCNN) was proposed to predict the spatial-temporal temperature field. A numerical database of subway tunnel fires was constructed based on the computational fluid dynamics (CFD) method. Considering different fire locations, heat release rates (HRRs), and ventilation velocities, 135 scenarios were conducted to collect fire data. An adaptive up-projection unit (AUPU) and axial attention fusion block were proposed to improve the TCNN. The results showed that, with the inputs of fire location, HRR, and ventilation velocity, the surrogate model can quickly identify given features and reproduce the spatial-temporal temperature field with 94 % accuracy within 0.034 s. The proposed deep learning surrogate model can aid emergency control and rescue efforts in the context of subway tunnel fires.