JEI thanks the reviewers who served the journal in 2024.

Uncalibrated photometric stereo is a formidable challenge task in the field of 3D vision, aiming to reconstruct the surface normal of an object when its shape, material reflectance, and light conditions are all unknown. At present, it remains difficult to address when dealing with more general materials (e.g., anisotropy) with complex reflectance and objects with significant shadows. In addition, the presence of generalized bas-relief ambiguity, which refers to the inherent ambiguity between shape and light, further compounds the challenges of uncalibrated photometric stereo. To overcome these limitations, we propose a lightweight unsupervised neural inverse rendering architecture, called General Material Shadow-Neural Inverse Rendering (GMS-NIR), which can effectively solve the uncalibrated photometric stereo problem by combining a learnable general material reflectance model and a shadow rendering model. We design an optimization strategy that allows GMS-NIR to jointly optimize the surface normal, light direction, and light intensity in an unsupervised manner by utilizing backpropagation to minimize rendering errors. Thorough experiments with diverse real-world datasets affirm the superior performance of our approach compared with alternative uncalibrated photometric stereo methods. GMS-NIR’s ability to handle a variety of materials and complex object shapes while accurately reconstructing surface normal makes it a promising advancement in the fields of computer vision and 3D surface reconstruction.

Gait recognition is a biometric technology that distinguishes individuals by analyzing their walking patterns and holds significant potential for development. Current methods primarily focus on extracting gait features either from the overall appearance or specific local regions. However, they often overlook the partitioning of local regions based on the individual’s body structure, as well as the weighted relationships between global and local features. We propose a Weighted Global-Local Feature Fusion Module to partition local features according to human body parts and adaptively integrate global and local gait features. This approach facilitates fine-grained learning of part-level local features and enhances the discriminative representation of gait features. Furthermore, we employ an Attention-based Multiscale Temporal Aggregation operation to adaptively fuse motion features from different time scales, preserving crucial spatio-temporal information while reducing the length of the time series. The average Rank-1 accuracy in CASIA-B and OUMVLP datasets is 93.0% and 90.7%, respectively. The experimental results demonstrate that our method achieves satisfactory recognition performance, indicating its potential for advancing gait recognition technology.

We propose a learned inverted residual cascaded group small target detection transformer (LIS-DETR) model, a novel approach for small object detection in autonomous driving. This model has a unique backbone network, the basic inverted residual cascaded group mobile block, which enhances feature representation and reduces computational redundancy. A dedicated small detection layer is integrated to improve small object detection specifically. In addition, an adaptive learned positional encoding transformer layer is incorporated to strengthen global contextual relationships, and the designed inner-SIoU loss function further accelerates convergence speed. Experimental results show a 3.1% increase in mAP50 accuracy on VisDrone datasets and a 1.9% improvement on processed SODA10m datasets compared with baseline methods. These advances demonstrate the LIS-DETR model’s strong generalization ability and the significant potential to enhance the efficacy of autonomous driving systems.

High-density printed circuit boards (PCBs) (with a minimum line width of 0.075 mm) display random and irregular defect patterns due to complex manufacturing processes, challenging their reliability. Deep learning-based visual detection methods show high accuracy but depend on the scale and diversity of datasets. The absence of large, well-annotated PCB defect datasets limits these methods in industrial applications. This paper introduces and open-sources an industrially applicable dataset, PCB-BB, which features large-scale, diverse, and high-fidelity data. In our proposed Zero-Shot Defect Feature Optimizer method, we first reconstruct the structure and color features of the learning image and encode these features into a unified codebook. Subsequently, by performing latent space codebook confusion and query decoding at the defect’s corresponding locations, we optimize the morphology and color characteristics of the image surface defects. Detection networks trained with optimized data show significant improvements: a 4.7% increase in F1 score, 2.7% in mAP@50-95, 2.5% in mAP@50, and 10.0% in recall. Our approach addresses key challenges such as scale mismatch and detection efficiency in practical deep learning applications. Enhancing visual detection for complex defects in industrial settings is crucial for advancing the intelligence of detection systems. All resources will be available on GitHub ( https://github.com/OrigArith/PCB-BB-ZSDF-OPT).

In recent years, RGB-T salient object detection (SOD) technology has attracted increasing attention. By incorporating thermal images, it can identify salient objects under challenging scenes such as low illumination. However, due to the inherent modality differences between RGB and thermal images, the interaction and fusion of multi-modal features become a critical challenge. To address this challenge, we propose a novel asymmetric light-aware progressive decoding network (ALPD-Net) for RGB-T SOD. Specifically, we develop an asymmetric light-aware interaction (ALI) module that facilitates effective interaction between the two modalities in an asymmetric manner, reducing interference information. In addition, we propose a Channel-Space Feature Fusion module to select and fuse information from different modalities in both channel and spatial dimensions. Finally, we design a phased progressive decoding strategy that divides the decoding process into two stages, gradually refining features to generate high-quality saliency maps. Extensive experiments conducted on three publicly available RGB-T SOD datasets demonstrate that the proposed ALPD-Net achieves outstanding performance against the state-of-the-art RGB-T SOD methods.

Aiming at the non-uniformity response problems of large array gazing remote sensing cameras due to factors such as production process and working mode, we proposed a non-uniformity correction method based on an improved total variation model. First, an improved total variation model with norm fidelity is constructed based on the distortion model of the imaging system. Next, the improved total variation model is iteratively solved by the Split Bregman splitting method. Then, in the iterative optimization process of the total variation model, the wavelet coefficients of the non-uniformity correction results of the improved total variation model are reconstructed using wavelet transform, and the high-frequency wavelet coefficients affecting the image quality are processed. Finally, the optimal iteration result obtained after wavelet reconstruction is the final correction result, and the non-uniformity correction of the remote sensing image is completed. The experimental results show that the non-uniformity correction effect of the proposed method is superior to other comparison methods. After correction, the non-uniformity coefficient of the image is reduced by more than three times, and the signal-to-noise ratio of the image is increased by more than 20 times, which effectively eliminates the non-uniformity response in the image and preserves the detail information of the image.

In low-light conditions, object detection algorithms suffer from reduced accuracy due to factors such as noise and insufficient information. Current solutions often involve a two-stage process: first, improving image illumination and then performing object detection. However, this method has limitations as these networks work independently. To address this, we propose a parallel object detection algorithm for low-light environments. Our approach simultaneously encodes image features using both an illumination enhancement network and an object detection network. This innovative design allows these networks to adapt to each other, improving feature adaptability for object detection. We enhance adaptive learning efficiency by introducing a novel mutual feedback mechanism, which dynamically adjusts the learning weights of the two networks, thereby enhancing the network’s capacity to encode object-related information in low-light conditions. Experiments were conducted on both real-world and synthetic datasets. On the real-world dataset, the proposed method outperformed the original object detection network, achieving improvements of 4.76% in mAP@0.5, 12.12% in mAP@0.5:0.95, and 8.4% in F1-score. On the synthetic dataset, the method demonstrated gains of 9.67%, 9.75%, and 10.6% in mAP@0.5, mAP@0.5:0.95, and F1-score, respectively. These experimental results indicate that the proposed approach significantly enhances the performance of object detection algorithms under low-light conditions.

The growing demand for immersive experiences has significantly influenced research in the quality assessment of light field images (LFIs). However, LFIs are susceptible to distortion during encoding, transmission, and compression, making accurate distortion measurement a current concern. In the full-reference (FR) quality assessment of LFIs, the discrepancy between the distorted and reference images is typically learned for network acquisition, overlooking the potential to enhance learning efficiency and accuracy by incorporating error map input. To address this, we propose a framework for the quality evaluation of FR LFIs based on multi-feature interactive fusion. The framework comprises three components: feature encoding network, spatial angle interactive fusion network, and score regression network, aimed at obtaining the quality score of LFIs. The feature encoding network leverages different extraction networks to capture spatial, angular, and error information, enabling the network to focus on key areas. In the spatial angle interactive fusion network, the feature fusion network integrates features from different networks, enriching and unifying information, the spatial angle interactive network further extracts distortion information from the enriched feature maps. The resulting framework demonstrates strong subjective and objective agreement in the quality assessment of LFIs, offering theoretical and practical implications for algorithm optimization and application.

In recent years, vision transformers (ViTs) have made significant breakthroughs in computer vision and have demonstrated great potential in large-scale models. However, the quantization methods for convolutional neural network models do not perform well on ViTs models, leading to a significant decrease in accuracy when applied to ViTs models. We extend the quantization parameter optimization method based on the Hessian matrix and apply it to the quantization of the LayerNorm module in ViT models. This approach reduces the impact of quantization on task accuracy for the LayerNorm module and enables more comprehensive quantization of ViT models. To achieve fast quantization of ViTs models, we propose a quantization framework specifically designed for ViTs models: Hessian matrix–aware post-training quantization for vision transformers (HAPTQ). The experimental results on various models and datasets demonstrate that our HAPTQ method, after quantizing the LayerNorm module of various ViT models, can achieve lossless quantization (with an accuracy drop of less than 1%) in ImageNet classification tasks. Specifically, the HAPTQ method achieves 85.81% top-1 accuracy on the ViT-L model.

Aiming at problems of key target inspection perception caused by harsh factors, such as high dust and fog, uneven illumination, mixed arrangement, and cross-scale variation of personnel-equipment elements in coal mine complex operation scenes, an inspection perception method of coal mining and excavating face based on machine vision is proposed. First, taking the YOLOv5s algorithm as the baseline model, the C3-DRES feature extraction module incorporating the dual attention mechanism and dual residual structure, the cross-layer dense connection and cross-scale connection (CDS-FFN) feature fusion network based on cross-layer dense connection and cross-scale connection, and the task-specific context decoupling head are designed to construct the coal mine complex inspection scene key objective detection network (CMCIS-Net). Second, based on the coal mine complex scene data set, the performance verification experiment of optimization strategies, ablation experiment, and comparison experiment are carried out, and the CMCIS-Net is deployed to the visual perception terminal platform of the inspection robot, to test the detection performance of the inspection robot on equipment and personnel. The experimental results show that in the complex operation scenes of coal mine, the three optimization strategies of C3-DRES, CDS-FFN, and task-specific context decoupling head effectively improve the target feature region focus and feature information fusion ability of the network and reduce the classification and positioning loss, to increase the detection accuracy of the network by 1.4%, 1.6%, and 1.9%, respectively; the detection accuracy of CMCIS-Net reaches 95.0%, which is 4.9% higher than that of the YOLOv5s algorithm, and compared with the eight mainstream target detection algorithms of SSD–YOLOv9c, CMCIS-Net has the best detection performance; on the vision control board, the real-time detection speed and inference time of CMCIS-Net reach 36.3 frames/s and 26.3 ms, and the detection accuracy of all targets is above 90.0%, which can realize the stable identification and accurate positioning of equipment and personnel.

Anomaly detection has various applications in the industrial field. However, most current studies in this field are idealized such that they can only address anomaly detection in simple scenarios. For instance, most studies assumed that all detection objects are centered and typically have a homogeneous monochromatic background, with little spatial position change across different images. Such an assumption makes it difficult to apply the research findings to complex industrial production scenarios. To address the limitations of existing idealized approaches, we propose a frequency-guided image reconstruction method with multi-level feature combination (MLFC) and multi-scale feature fusion (MSFF) for anomaly detection. The overview of the proposed method is as follows: First, we employ a mixed anomaly generation strategy to generate anomalous samples for training. Second, before feeding the images into the reconstruction network, we introduce a frequency selection module. This module assigns different frequency weights to different frequency components, to suppress the tendency toward identity mapping of the anomalies of the network. Finally, we design a reconstruction network based on U-Net by employing three MLFC modules and one MSFF module. The MLFC modules in the skip-connections adaptively fuse features from multiple encoder levels using channel attention. The MSFF module thereafter fuses the multi-scale features from the decoder to enhance the representational capability of the decoder output features. The experimental results demonstrate that even with a simple U-Net as the reconstruction network, the proposed method achieves state-of-the-art performance on the metal parts defect detection dataset. Specifically, it reaches an image-level area under the receiver operating curve (AUROC) of 94.91% and a pixel-level AUROC of 97.88%, setting a new benchmark.

One of the principal causes of mining accidents is the unsafe conduct of miners. The monitoring of miners’ unsafe actions and the timely implementation of corrective measures can effectively prevent a considerable number of preventable accidents. The SlowFast network represents a representative network for video action recognition. In comparison with conventional two-stream networks, the SlowFast networks avoid the extraction of optical flow information. Nevertheless, the quality of the imaging of coal mine images is inferior to that of regular images, exhibiting deficiencies such as inadequate illumination and the loss of fine details. Furthermore, the original network was incapable of effectively integrating horizontal information, resulting in a further reduction in accuracy when testing for coal miners. To address these issues, a shuffle attention SlowFast network for video action recognition has been proposed. A lightweight yet efficient shuffle attention module is employed to accurately focus on all the relevant elements of the lateral connections while also focusing action information through the attention mechanism. To address the issue of data imbalance, a linear combination of polynomial functions, termed PolyLoss, is employed. The proposed architecture was evaluated on three datasets: the public UCF-101 dataset, the public HMDB-51 dataset, and a self-constructed dataset of miner actions. The experimental results demonstrated that the proposed method exhibited superior performance compared with existing state-of-the-art methods for both regular images and images of underground mines.

Crane defect detection traditionally relies on manual inspection and image processing techniques, which struggle with complex defect types and subtle features. Although deep learning has improved detection performance, challenges with insufficient detection capability persist. We propose a defect recognition and localization method for crane components using the YOLOv5s-CUA fusion enhancement algorithm. A dataset specifically for floating crane defects is first established. The YOLOv5s model is then enhanced by replacing nearest-neighbor interpolation up-sampling with the content-aware reassembly of features operator to improve feature map completeness, integrating the convolutional block attention module with the C3 module to boost important feature channel extraction, and incorporating the adaptive structure feature fusion module to enhance adaptive fusion capabilities. Experimental results demonstrate that the improved YOLOv5s model increases detection accuracy by 2% and recall rate by 3.7% compared with the original YOLOv5s model, significantly reducing leakage and misdetection. This advancement enables rapid identification of defects in floating crane components and provides robust technical support for practical applications.

Traditional partial least squares regression typically takes vectorized data into account. The process of vectorizing images may disrupt the inherent structural information of the data, leading to the curse of dimensionality. To address this problem, we propose a model called two-dimensional partial least squares regression with manifold optimization–based Gaussian filter, first introducing the Gaussian filter to remove noise in the data and smooth the images. Subsequently, during the establishment of the model, features are extracted using the two-dimensional principal component analysis approach to preserve the original structure of the data. Ultimately, the objective function of the model is constructed and maximized. A detailed derivation process for the objective function is conducted. Given that the projection matrices adhere to the column orthogonal constraint, a manifold optimization approach is utilized to search for the optimal projection matrices on Stiefel manifold space. Experimental results on various real datasets demonstrate that our proposed approach has superior adaptability and robustness compared with other representative approaches.

Underwater target detection plays an indispensable role in the fields of marine exploration, resource exploitation, and marine protection. However, complex underwater environments, such as insufficient light, turbid water, and target occlusion, place high demands on the accuracy and effectiveness of target detection. To address the above problems, we propose the ESCL-YOLO underwater target detection algorithm based on improved YOLOv8. First, we design an efficient shared detail-enhanced convolutional detection head, which reduces the number of model parameters while improving the performance of the detection head in localization and classification as well as the ability of detail capture. To enhance the ability of model perception and extraction of fine features, space-to-depth convolution is introduced to replace part of the ordinary convolution in YOLOv8. On this basis, the Content-aware Reassembly of Features upsampling operator is introduced to improve the content perception ability of the model while increasing the sensory field of the model. Finally, a large separable kernel attention is introduced into the spatial pyramid pooling layer of the model, which allows the model to accurately capture the feature information of interest at different scales, further improving the accuracy of the model. Through extensive experiments and analysis on the 2020 National Underwater Robotics Professional Competition dataset and the real-world underwater object detection dataset, the precision, recall, and mean average precision (mAP50) of the algorithm reached 84.6%, 76.9%, and 84.2% and 85.9%, 79.9%, and 86.4%, respectively. Compared with the original model, the mean average precision (mAP50) is improved by 1.6% and 1.3%.

We propose an adaptive enhancement algorithm for low-illumination images based on improved multi-scale Retinex with color restoration (MSRCR) to address the problems of low contrast, poor visual quality, and blurred edge information of images under complex illumination conditions in an automotive welding workshop. First, the reciprocal of the just noticeable difference is added to the MSRCR algorithm as an adjustment factor for the illumination component to obtain an image with clear edge information; then, the adaptive contrast enhancement algorithm is utilized to obtain an image with high contrast and high luminance; finally, two images are fused by Laplace pyramid fusion in proportion to the luminance mean value and linearly stretched. This algorithm can effectively solve the color bias problem of the MSRCR algorithm, highlight the detailed information of the edge of the welded stud to be measured, and retain the brightness enhancement effect.

Deep learning algorithms are extensively employed in the detection of facial forgery. The foundation of these methods relies on the utilization of a large number of frames for learning purposes. In contrast to other state-of-the-art methods that require numerous consecutive or uniformly sampled frames as network input, we propose a key frame strategy to maximize inter-frame dissimilarities. We utilize the weighted energy of optical flow to identify key frames in the video, thereby enhancing the quality of features through CNN-LSTM training. Evaluations conducted on FaceForensics++, Celeb-DF, and DFDC datasets demonstrate that the proposed method achieves superior or comparable performance in terms of within and cross dataset testing.

Text-to-image person re-identification (TIReID) is a significant challenge in the cross-modal community, focused on retrieving individuals based on textual queries. The primary obstacle is effectively mapping visual and textual modalities into a shared latent space, a problem inadequately addressed by previous methods that relied on separately pre-trained unimodal models, which often lack the necessary alignment capabilities. Recently, contrastive language-image pre-training (CLIP) has emerged as a versatile large-scale cross-modal visual-language pre-training model, excelling in various cross-modal downstream tasks due to its powerful semantic learning capabilities. CLIP has successfully addressed the need for manual alignment of body part features required by earlier methods. However, in TIReID, integrating CLIP with ReID presents notable alignment issues. It struggles with filtering task-specific irrelevant information, leading to redundancy and interference. In addition, CLIP lacks effective internal alignment, such as misaligned body parts and semantic misalignment. Finally, the joint loss function integrates identification loss, image-text contrastive loss, and mask-based unsupervised training to enhance feature alignment. This new loss structure effectively reduces the risk of over-alignment, ensuring a more balanced training process. Experimental results on CUHK-PEDES benchmarks demonstrate cross-modal latent feature alignment’s effectiveness, surpassing state-of-the-art methods with improvements in rank 1 accuracy by 1.54%.

Six-degrees-of-freedom (6-DoF) object pose accurate estimation is an important visual task. However, due to complex application scenarios such as complicated target shapes and mutual occlusion, etc., traditional feature-matching methods result in low-pose estimation accuracy. Deep learning methods use multi-source information about the target and employ dense fusion prediction networks to estimate the target pose, providing high accuracy and robustness. However, the accuracy of target pose estimation depends on the input data accuracy of the prediction network. Therefore, we proposed an improved instance segmentation network SOLOv2 to identify and segment targets in complex scenarios based on the RGB-D data of targets, extracting multi-source information of the target such as color, depth, and mask. Then, we used an improved dense fusion pose estimation network SOLO-Dense to fuse the extracted multi-source information to achieve object 6-DoF pose estimation. Experiments showed that the average precision (AP) of the improved segmentation network increased by 5.1% compared with the original network. The results of our pose estimation method showed an excellent level of evaluation of the LineMOD and YCB-Video datasets. This method accurately segments targets and estimates their pose in cluttered scenarios, demonstrating the effectiveness of the proposed approach.

Underwater object detection is an important computer vision task that has been widely used in marine life identification and tracking. However, there are a series of challenges such as low-contrast condition, occlusion condition, unbalanced light condition, and small dense objects in underwater object detection. Attention mechanism has been proven powerful in feature extraction. However, attention mechanisms usually divide image into fixed patches. This methods lead to the splitting up of continuous structures, which hinders the use of similar information in other areas to enhance image details. In fact, using graph structure is more flexible and effective for visual perception, because it can capture complex target relation and context information effectively. Thus, we apply graph attention mechanisms to irregular patches and propose an irregular-patch graph attention network (IGA-Net). First, the superpixel segmentation method is used to segment the image to reduce noise. Second, the global graph and local graph are constructed to obtain internal structures. Finally, to handle occlusion and small objects, a distinctive feature three-way handshake (F3H) module is proposed to fuse information from global and local graph. To demonstrate the effectiveness of the proposed method, we conduct comprehensive evaluations on five challenging underwater datasets UTDAC2020, RUOD, Brackish, TrashCan, and WPBB. Experimental results demonstrate that the proposed IGA-Net achieves superior performance on five challenging underwater datasets.

In recent times, within the domain of the unmanned aerial vehicle (UAV) domain, there has been extensive utilization of discriminative correlation filter (DCF) techniques for object tracking, largely attributed to their effective performance. However, conventional DCF-based tracking methods rely solely on historical information to train the filters. Consequently, when encountering specific UAV tracking challenges, such as viewpoint changes, these trackers may be disturbed by new environments, which may result in tracking failure. In addition, template degradation due to the boundary effect can further degrade the performance of the filters. To overcome the aforementioned difficulties, we present a correlation filter approach that leverages smoothing prediction and spatio-temporal regularization. First, we employ single exponential smoothing prediction to provide an approximate estimation of the object’s position in upcoming frames. Second, a historical filtering model is formulated, accompanied by the incorporation of a spatio-temporal regularization component. Finally, we utilize the future frame context information and the historical filter model to train the filter for tracking. The experimental results on the DTB70, UAV123_10fps, UAVTrack112, and UAVDT benchmarks demonstrate the efficacy of our approach and indicate that the operation speed of 68 frames per second can be achieved on a central processing unit (CPU).

Depth images captured by commercial red green blue-depth (RGB-D) cameras commonly suffer from missing regions, which severely hinders their applicability. To effectively restore missing depth values, especially for large holes, we propose a depth image restoration based on uncertainty metrics (DRUM) for hole-filling. Specifically, pseudo-background and information entropy are proposed to quantify the uncertainty for determining the filling priority and identifying pixels needing restoration, which effectively mitigates error accumulation. Subsequently, color and gradient clues are fully exploited to guide depth value prediction, remarkably enhancing accuracy. Then, ablation studies validate the efficacy of the proposed innovations. Comparative experiments on public benchmarks demonstrate DRUM’s superior performance over prevailing methods on diverse hole patterns. Finally, a fast implementation called Fast-DRUM is devised through multi-threading and selective updating. Comparative evaluations of processing time demonstrate its high efficiency. We provide new insights into reliable depth completion for practical applications.

In complex road scenarios, issues such as low-detection accuracy of positive and negative obstacles on the road surface and weak generalization ability of detection models are prevalent. We propose a high-precision detection network for positive and negative road obstacles, named multiscale diffusion and mix-attention transformer (MSD-MixFormer), which is based on multiscale fusion and diffusion mix-attention mechanisms. This network is designed based on the transformer architecture with a window adjustment mechanism to capture multi-scale features of obstacles, thereby addressing the issue of confusion between obstacle foreground and background. In the shallow and deep stages of the encoder module, dilation self-attention and diffusion self-attention are introduced respectively to enhance the feature fusion capability of the network. In the decoder module, mix-attention operations are incorporated to improve the segmentation accuracy of detailed features, addressing the problem of detail feature loss. To improve the generalization ability of the model, we collect and extend the VRE-NPO dataset with images from actual complex road scenarios. Experiments show that the MSD-MixFormer model converges faster and more stably during training compared with other higher-level models in the same series. In testing, MSD-MixFormer achieves higher detection accuracy for positive and negative obstacles, as well as for objects of varying sizes, across various complex road scenarios.

We focus on the complexity and multi-scale importance of semantic segmentation of remote sensing images, a fundamental task in earth science research. We propose an architecture, TransUNetFormer, that is potent such as U-Net designed that extensively integrates convolutional neural network (CNN) + transformer fusion in both the encoder and decoder; this fusion emphasizes the significance of global contextual information and local feature details. TransUNetFormer achieves superior generalization for remote sensing image segmentation, particularly in capturing multi-scale features within its encoder-decoder architecture. The encoder incorporates design principles inspired by TransUNet, leveraging a CNN + transformer component for an efficient hybrid. In addition, a CNN + transformer hybrid block in the decoder, DP-hybrid, efficiently captures rich global-local features at each upsampling step. We introduce a fusion-concatenation module to dynamically generate weights during the interaction between the encoder and decoder, facilitating feature map fusion. Finally, an efficient feature refinement segmentation head is devised to purify shallow-stage encoder features and optimize the most profound global-local features in the decoder for fusion output. Experimental results on two widely used datasets, ISPRS Potsdam and LoveDA Urban, show the effectiveness and potential of TransUNetFormer. To our knowledge, this is the first hybrid CNN + transformer network specifically designed for remote sensing image segmentation.