Decoder Structure Research Articles

Multiple Unmanned Aerial Vehicle (multi-UAV) Reconnaissance and Search with Limited Communication Range Using Semantic Episodic Memory in Reinforcement Learning

Unmanned Aerial Vehicles (UAVs) have garnered widespread attention in reconnaissance and search operations due to their low cost and high flexibility. However, when multiple UAVs (multi-UAV) collaborate on these tasks, a limited communication range can restrict their efficiency. This paper investigates the problem of multi-UAV collaborative reconnaissance and search for static targets with a limited communication range (MCRS-LCR). To address communication limitations, we designed a communication and information fusion model based on belief maps and modeled MCRS-LCR as a multi-objective optimization problem. We further reformulated this problem as a decentralized partially observable Markov decision process (Dec-POMDP). We introduced episodic memory into the reinforcement learning framework, proposing the CNN-Semantic Episodic Memory Utilization (CNN-SEMU) algorithm. Specifically, CNN-SEMU uses an encoder–decoder structure with a CNN to learn state embedding patterns influenced by the highest returns. It extracts semantic features from the high-dimensional map state space to construct a smoother memory embedding space, ultimately enhancing reinforcement learning performance by recalling the highest returns of historical states. Extensive simulation experiments demonstrate that in reconnaissance and search tasks of various scales, CNN-SEMU surpasses state-of-the-art multi-agent reinforcement learning methods in episodic rewards, search efficiency, and collision frequency.

ESFPNet: Efficient Stage-Wise Feature Pyramid on Mix Transformer for Deep Learning-Based Cancer Analysis in Endoscopic Video.

For patients at risk of developing either lung cancer or colorectal cancer, the identification of suspect lesions in endoscopic video is an important procedure. The physician performs an endoscopic exam by navigating an endoscope through the organ of interest, be it the lungs or intestinal tract, and performs a visual inspection of the endoscopic video stream to identify lesions. Unfortunately, this entails a tedious, error-prone search over a lengthy video sequence. We propose a deep learning architecture that enables the real-time detection and segmentation of lesion regions from endoscopic video, with our experiments focused on autofluorescence bronchoscopy (AFB) for the lungs and colonoscopy for the intestinal tract. Our architecture, dubbed ESFPNet, draws on a pretrained Mix Transformer (MiT) encoder and a decoder structure that incorporates a new Efficient Stage-Wise Feature Pyramid (ESFP) to promote accurate lesion segmentation. In comparison to existing deep learning models, the ESFPNet model gave superior lesion segmentation performance for an AFB dataset. It also produced superior segmentation results for three widely used public colonoscopy databases and nearly the best results for two other public colonoscopy databases. In addition, the lightweight ESFPNet architecture requires fewer model parameters and less computation than other competing models, enabling the real-time analysis of input video frames. Overall, these studies point to the combined superior analysis performance and architectural efficiency of the ESFPNet for endoscopic video analysis. Lastly, additional experiments with the public colonoscopy databases demonstrate the learning ability and generalizability of ESFPNet, implying that the model could be effective for region segmentation in other domains.

MAS-Net: Multi-Attention Hybrid Network for Superpixel Segmentation

Superpixels, as essential mid-level image representations, have been widely used in computer vision due to their computational efficiency and redundant compression. Compared with traditional superpixel methods, superpixel algorithms based on deep learning frameworks demonstrate significant advantages in segmentation accuracy. However, existing deep learning-based superpixel algorithms suffer from a loss of details due to convolution and upsampling operations in their encoder–decoder structure, which weakens their semantic detection capabilities. To overcome these limitations, we propose a novel superpixel segmentation network based on a multi-attention hybrid network (MAS-Net). MAS-Net is still based on an efficient symmetric encoder–decoder architecture. First, utilizing residual structure based on a parameter-free attention module at the feature encoding stage enhanced the capture of fine-grained features. Second, adoption of a global semantic fusion self-attention module was used at the feature selection stage to reconstruct the feature map. Finally, fusing the channel with the spatial attention mechanism at the feature-decoding stage was undertaken to obtain superpixel segmentation results with enhanced boundary adherence. Experimental results on real-world image datasets demonstrated that the proposed method achieved competitive results in terms of visual quality and metrics, such as ASA and BR-BP, compared with the state-of-the-art approaches.

DAT-Net: Filling of missing temperature values of meteorological stations by data augmentation attention neural network

Abstract For a long time, filling in the missing temperature data from meteorological stations has been crucial for researchers in analyzing climate variation cases. In previous studies, people have attempted to solve this problem by using interpolation and deep learning methods. Through extensive case studies, it is observed that the data utilization rate of convolutional neural networks based on PConv is low at a high missing rate, which will result in the poor filling performance of each model at a high missing rate. To solve these problems, a Data Augmentation Attention Neural Network (DAT-Net) is presented. DAT Net uses encoder and decoder structures, which include a data augmentation training mechanism (DAM) to enhance model training. In addition, a time encoder (TED) has been developed to assist the model in learning the temporal dependencies of the data. To evaluate DAT-Net, 75% and 85% of experiments were performed, while comparisons were made with Linear, NLinear, DLinear, PatchTST, and GSTA-Net. The results showed that when the missing rate was 75%, DAT-Net decreased by 55.22%, 55.05%, 55.18%, 28.73%, and 12.35% on MAE and 54.08%, 53.88%, 54.08%, 35.48% and, 14.51% on RMSE, R 2 increased by 3.80%, 3,75%, 3.68%, 0.55%, and 0.27%, respectively.

Enhancing deep reinforcement learning for scale flexibility in real-time strategy games

Real-time strategy (RTS) games present a unique challenge for AI agents due to the combination of several fundamental AI problems. While Deep Reinforcement Learning (DRL) has shown promise in the development of autonomous agents for the genre, existing architectures often struggle with games featuring maps of varying dimensions. This limitation hinders the agent’s ability to generalize its learned strategies across different scenarios. This paper proposes a novel approach that overcomes this problem by incorporating Spatial Pyramid Pooling (SPP) within a DRL framework. We leverage the GridNet architecture’s encoder–decoder structure and integrate an SPP layer into the critic network of the Proximal Policy Optimization (PPO) algorithm. This SPP layer dynamically generates a standardized representation of the game state, regardless of the initial observation size. This allows the agent to effectively adapt its decision-making process to any map configuration. Our evaluations demonstrate that the proposed method significantly enhances the model’s flexibility and efficiency in training agents for various RTS game scenarios, albeit with some discernible limitations when applied to very small maps. This approach paves the way for more robust and adaptable AI agents capable of excelling in sequential decision problems with variable-size observations.

Arbitrary style transfer via multi-feature correlation

Recent research in arbitrary style transfer has highlighted challenges in maintaining the balance between content structure and style patterns. Moreover, the improper application of style patterns onto the content image often results in suboptimal quality. In this paper, a novel style transfer network, called MCNet, is proposed. It is based on multi-feature correlations. To better explore the intrinsic relationship between the style image and the content image and to transfer the most suitable style onto the content image, a novel Global Style-Attentional Transfer Module, named GSATM, is introduced in this work. GSATM comprises two parts: Forward Adaptive Style Transformation (FAST) and Delayed Style Transformation (DST). The former analyzes the relationship between style and content features and fine-tunes the style features, whereas the latter transfers the content features based on the fine-tuned style features. Moreover, a new encoding and decoding structure is designed to effectively handle the output of GSATM. Extensive quantitative and qualitative experiments fully demonstrate the superiority of our algorithm. Project page: https://github.com/XiangJinCherry/MCNet.

Research on Multi-Step Fruit Color Prediction Model of Tomato in Solar Greenhouse Based on Time Series Data

Color change is the most obvious characteristic of the tomato ripening stage and an important indicator of the tomato ripening condition, which directly affects the commodity value of tomato. To visualize the color change of tomato fruit during the mature stage, this paper proposes a gated recurrent unit network with an encoder–decoder structure. This structure dynamically simulates the growth and development of tomatoes using time-dependent lines, incorporating real-time information such as tomato color and shape. Firstly, the .json file was converted into a mask.png file, the tomato mask was extracted, and the tomato was separated from the complex background environment, thus successfully constructing the tomato growth and development dataset. The experimental results showed that for the gated recurrent unit network with the encoder–decoder structure proposed, when the hidden layer number was 1 and hidden layer number was 512, a high consistency and similarity between the model predicted image sequence and the actual growth and development image sequence was realized, and the structural similarity index measure was 0.746. It was proved that when the average temperature was 24.93 °C, the average soil temperature was 24.06 °C, and the average light intensity was 11.26 Klux, the environment was the most suitable for tomato growth. The environmental data-driven tomato growth model was constructed to explore the growth status of tomato under different environmental conditions, and thus, to understand the growth status of tomato in time. This study provides a theoretical foundation for determining the optimal greenhouse environmental conditions to achieve tomato maturity and it offers recommendations for investigating the growth cycle of tomatoes, as well as technical assistance for standardized cultivation in solar greenhouses.

CrowdUNet: Segmentation assisted U-shaped crowd counting network

With the end of the COVID-19 pandemic, the number of pedestrians in various public places has increased dramatically. Estimating the size and density distribution of crowds accurately from images is essential for public safety. At present, there are still many factors that limit the accuracy of dense crowd counting, such as perspective distortion, background clutter and heavy occlusion. To be capable of accurately estimating the crowd size in RGB images, we propose a crowd counting network called CrowdUNet, which is assisted by a segmentation task. It applies the segmentation results to the crowd counts, making the network more focused on the prediction of foreground regions. We combine Swin Transformer Block and CNN to build a Swin Transformer Convolution(STC) module to extract deep semantic features. We analyze the characteristics of crowd images and propose a novel decoder structure called Coordinate Decoder(CD), which better aggregates low and high level features and improve the robustness of the network. In order to obtain accurate regression results, we also propose a regression head with multi-scale receptive fields, which is called Spatial Pyramid Convolution (SPC). Extensive experiments on four challenging crowd counting datasets namely ShanghaiTech A, ShanghaiTech B, UCF p=CC 50, and UCF-QNRF have validated the proposed method.

A deep learning model for predicting the state of energy in lithium-ion batteries based on magnetic field effects

The state of energy (SOE) is one of the most critical state indicators in battery management systems. However, its nonlinear characteristics present significant challenges in obtaining accurate SOE. Especially when applying different magnetic field strengths to perform battery charging and discharging tests, the change in battery energy becomes more complex due to the influence of the magnetization effect. In this paper, a deep learning network, combining an improved Informer and long short-term memory network (LSTM), was developed to estimate the SOE of lithium-ion batteries under different magnetic fields. First, we improve the decoder structure by adding a convolutional module using residual connections with trainable weight parameters to capture hidden states with more details.The improved decoder does not require label history information for decoding, which improves the generalization ability of the model. Finally, the output of the Informer network is a higher-dimensional hidden feature that is input into the LSTM network layer to output the SOE prediction value, which improves the original Informer network's ability to integrate sequences. Experiments with magnetic field and public datasets show the improved Informer-LSTM network achieves 0.31 % MAE, 0.42 % RMSE, and 1.79 % maximum error in SOE estimation, outperforming others in short sequence predictions.

A Benchmark for Morphological Segmentation in Uyghur and Kazakh

Morphological segmentation and stemming are foundational tasks in natural language processing. They have become effective ways to alleviate data sparsity in agglutinative languages because of the nature of agglutinative language word formation. Uyghur and Kazakh, as typical agglutinative languages, have made significant progress in morphological segmentation and stemming in recent years. However, the evaluation metrics used in previous work are character-level based, which may not comprehensively reflect the performance of models in morphological segmentation or stemming. Moreover, existing methods avoid manual feature extraction, but the model’s ability to learn features is inadequate in complex scenarios, and the correlation between different features has not been considered. Consequently, these models lack representation in complex contexts, affecting their effective generalization in practical scenarios. To address these issues, this paper redefines the morphological-level evaluation metrics: F1-score and accuracy (ACC) for morphological segmentation and stemming tasks. In addition, two models are proposed for morpheme segmentation and stem extraction tasks: supervised model and unsupervised model. The supervised model learns character and contextual features simultaneously, then feature embeddings are input into a Transformer encoder to study the correlation between character and context embeddings. The last layer of the model uses a CRF or softmax layer to determine morphological boundaries. In unsupervised learning, an encoder–decoder structure introduces n-gram correlation assumptions and masked attention mechanisms, enhancing the correlation between characters within n-grams and reducing the impact of characters outside n-grams on boundaries. Finally, comprehensive comparative analyses of the performance of different models are conducted from various points of view. Experimental results demonstrate that: (1) The proposed evaluation method effectively reflects the differences in morphological segmentation and stemming for Uyghur and Kazakh; (2) Learning different features and their correlation can enhance the model’s generalization ability in complex contexts. The proposed models achieve state-of-the-art performance on Uyghur and Kazakh datasets.

Dual-branch deep cross-modal interaction network for semantic segmentation with thermal images

Semantic segmentation using RGB (Red-Green-Blue) images and thermal datas is an indispensable component of autonomous driving. The key to RGB-Thermal (RGB and Thermal) semantic segmentation is achieving the interaction and fusion of features between RGB and thermal images. Therefore, we propose a dual-branch deep cross-modal interaction network (DCIT) based on Encoder–Decoder structure. This framework consists of two parallel networks for feature extraction from RGB and Thermal data. Specifically, in each feature extraction stage of the Encoder, we design a Cross Feature Regulation Modules (CFRM) to align and correct modality specific features by reducing the inter-modality feature differences and eliminating intra-modality noise. Then, the modality features are aggregated through Cross Modal Feature Fusion Module (CMFFM) based on cross linear attention to capture global information from modality features. Finally, Adaptive Multi-Scale Cross-positional Fusion Module (AMCFM) utilizes the fused features to integrate consistent semantic information in the Decoder stage. Our framework can improve the interaction of cross modal features. Extensive experiments on urban scene datasets demonstrate that our proposed framework outperforms other RGB-Thermal semantic segmentation methods in terms of objective metrics and subjective visual assessments.

Representation Learning Based on Vision Transformer

In recent years, with the rapid development of information technology, the volume of image data has grown exponentially. However, these datasets typically contain a large amount of redundant information. To extract effective features and reduce redundancy from images, a representation learning method based on the Vision Transformer (ViT) has been proposed, and to our best knowledge, Transformer was first applied to zero-shot learning (ZSL). The method adopts a symmetric encoder–decoder structure, where the encoder incorporates Multi-Head Self-Attention (MSA) mechanism of ViT to reduce the dimensionality of image features, eliminate redundant information, and decrease computational burden. Consequently, it effectively extracts features, and the decoder is utilized for reconstructing image data. We evaluated the representation learning capability of the proposed method in various tasks, including data visualization, image reconstruction, face recognition, and ZSL. By comparing with state-of-the-art representation learning methods, the outstanding results obtained validate the effectiveness of this method in the field of representation learning.

Robot Grasp Detection with Loss-Guided Collaborative Attention Mechanism and Multi-Scale Feature Fusion

Grasp detection serves as the fundamental element for achieving successful grasping in robotic systems. The encoder–decoder structure has become widely adopted as the foundational architecture for grasp detection networks due to its inherent advantages of speed and accuracy. However, traditional network structures fail to effectively extract the essential features required for accurate grasping poses and neglect to eliminate the checkerboard artifacts caused by inversion convolution during decoding. Aiming at overcoming these challenges, we propose a novel generative grasp detection network (LGAR-Net2). A transposed convolution layer is employed to replace the bilinear interpolation layer in the decoder to remove the issue of uneven overlapping and consequently eliminate checkerboard artifacts. In addition, a loss-guided collaborative attention block (LGCA), which combines attention blocks with spatial pyramid blocks to enhance the attention to important regions of the image, is constructed to enhance the accuracy of information extraction. Validated on the Cornell public dataset using RGB images as the input, LGAR-Net2 achieves an accuracy of 97.7%, an improvement of 1.1% over the baseline network, and processes a single RGB image in just 15 ms.

From simple labels to semantic image segmentation: leveraging citizen science plant photographs for tree species mapping in drone imagery

Abstract. Knowledge of plant species distributions is essential for various application fields, such as nature conservation, agriculture, and forestry. Remote sensing data, especially high-resolution orthoimages from unoccupied aerial vehicles (UAVs), paired with novel pattern-recognition methods, such as convolutional neural networks (CNNs), enable accurate mapping (segmentation) of plant species. Training transferable pattern-recognition models for species segmentation across diverse landscapes and data characteristics typically requires extensive training data. Training data are usually derived from labor-intensive field surveys or visual interpretation of remote sensing images. Alternatively, pattern-recognition models could be trained more efficiently with plant photos and labels from citizen science platforms, which include millions of crowd-sourced smartphone photos and the corresponding species labels. However, these pairs of citizen-science-based photographs and simple species labels (one label for the entire image) cannot be used directly for training state-of-the-art segmentation models used for UAV image analysis, which require per-pixel labels for training (also called masks). Here, we overcome the limitation of simple labels of citizen science plant observations with a two-step approach. In the first step, we train CNN-based image classification models using the simple labels and apply them in a moving-window approach over UAV orthoimagery to create segmentation masks. In the second phase, these segmentation masks are used to train state-of-the-art CNN-based image segmentation models with an encoder–decoder structure. We tested the approach on UAV orthoimages acquired in summer and autumn at a test site comprising 10 temperate deciduous tree species in varying mixtures. Several tree species could be mapped with surprising accuracy (mean F1 score =0.47). In homogenous species assemblages, the accuracy increased considerably (mean F1 score =0.55). The results indicate that several tree species can be mapped without generating new training data and by only using preexisting knowledge from citizen science. Moreover, our analysis revealed that the variability in citizen science photographs, with respect to acquisition data and context, facilitates the generation of models that are transferable through the vegetation season. Thus, citizen science data may greatly advance our capacity to monitor hundreds of plant species and, thus, Earth's biodiversity across space and time.

Bidirectional interaction of CNN and Transformer for image inpainting

Image inpainting requires filling the corrupted regions with plausible contents and fitting seamlessly with the surroundings. Currently, CNN-dominated inpainting models still produce artifacts inconsistent with surroundings when facing large-scale corrupted regions. This is attributed to the locality and weight sharing of convolutions, which limits the ability of CNNs to learn the correlations between corrupted contents and distant contexts explicitly. In contrast, Transformer-dominated models are experts in modeling such long-range correlations, but they often deteriorate local details due to the lack of locality. A potential inpainting model should incorporate the advantages of CNN and Transformer. To do this, we propose an inpainting-oriented concurrent hybrid network, termed bidirectional interaction dual-stream network (BIDS-Net), which can sufficiently integrate the advantages of CNN and Transformer. Specifically, we adopt a CNN stream to capture rich local patterns to achieve detail reconstruction and refinement, and a Transformer stream to model long-range contextual correlations to fully borrow the relevant information within a global field of view. More importantly, we devise a bidirectional feature interaction (BFI) module based on a selective feature fusion method, which aims to enhance the locality of the transformer stream and the long-range correlation awareness of the CNN stream to realize full complementary advantages. In addition, we build both of the streams upon a hierarchical encoder–decoder structure to facilitate multi-scale context reasoning and improve computational efficiency. The comprehensive experimental results on three public datasets demonstrate the superiority of the proposed BIDS-Net in image inpainting task.

Multi‐scale feature aggregation network for single‐image dehazing

AbstractTransformer possesses a broader perceptual scope, while the Convolutional Neural Network (CNN) excels at capturing local information. In this paper, the authors propose the Multi‐Sclale Feature Aggregation Network (MSFA‐Net) for single‐image dehazing which is fused with the advantages of Transformer and CNN. Our MSFA‐Net is based on the encoder–decoder structure, and there are four main innovations. Firstly, the authors make some improvements to the original Swin Transformer to make it more effective for dehazing tasks, and the authors name it Spatial Information Aggregation Transformer (SIAT). The authors place the SIAT in both encoder and decoder of MSFA‐Net for feature extraction. The authors propose an upsampling module called Efficient Spatial Resolution Recovery (ESRR) which is placed in the decoder part. Compared to commonly used transposed convolutions, the authors’ ESRR module has fewer computational cost. Considering that the haze distribution is always uneven and the information from each channel is different, the authors introduce the Dynamic Multi‐Attention (DMA) module to provide pixel‐wise weights and channel‐wise weights for input features. The authors place the DMA module between the encoder and decoder parts. As the network depth increases, the spatial structural information from the high‐resolution layer tends to degrade. To deal with the problem, the authors propose the Multi‐Scale Feature Fusion (MSFF) module to recover missing spatial structural information. The authors place the MSFF module in both the encoder and decoder parts. Extensive experimental results show that the authors’ proposed dehazing network achieves state‐of‐the‐art dehazing performance with relatively low computational cost.

A multimodal stepwise-coordinating framework for pedestrian trajectory prediction

Pedestrian trajectory prediction from the first-person view has still been considered one of the challenging problems in automatic driving due to the difficulty of understanding and predicting pedestrian actions. Observing that pedestrian motion naturally contains the repetitive pattern of the gait cycle and global intention information, we design a Multimodal Stepwise-Coordinating Network, namely MSCN, to sufficiently leverage the underlying human motion properties. Specifically, we first design a multimodal spatial-frequency encoder, which encodes the periodicity of pedestrian motion with a frequency-domain enhanced Transformer and other visual information with a spatial-domain Transformer. Then, we propose a stepwise-coordinating decoder structure, which leverages both local and global information in sequence decoding through a two-stage decoding process. After generating a coarse sequence from the stepwise trajectory predictor, we design a coordinator to aggregate the corresponding representations used to generate the coarse sequence. Subsequently, the coordinator learns to output a refined sequence through a knowledge distillation process based on the aggregated representations. In this way, MSCN can adequately capture the representations of short-term motion behaviors, thus modeling better long-term sequence prediction. Extensive experiments show that the proposed model can achieve significant improvements over state-of-the-art approaches on the PIE and JAAD datasets by 16.1% and 16.4% respectively.

Dynamic Perception-Based Vehicle Trajectory Prediction Using a Memory-Enhanced Spatio-Temporal Graph Network

In the realm of intelligent transportation systems, accurately predicting vehicle trajectories is paramount for enhancing road safety and optimizing traffic flow management. Addressing the impacts of complex traffic environments and efficiently modeling the diverse behaviors of vehicles are the key challenges at present. To achieve precise prediction of vehicle trajectories, it is essential to fully consider the dynamic changes in traffic conditions and the long-term dependencies of time-series data. In response to these challenges, we propose the Memory-Enhanced Spatio-Temporal Graph Network (MESTGN), an innovative model that integrates a Spatio-Temporal Graph Convolutional Network (STGCN) with an attention-enhanced Long Short-Term Memory (LSTM)-based sequence to sequence (Seq2Seq) encoder–decoder structure. MESTGN utilizes STGCN to capture the complex spatial dependencies between vehicles and reflects the interactions within the traffic network through road traffic data and network topology, which significantly influences trajectory prediction. Additionally, the model focuses on historical vehicle trajectory data points using an attention-weighted mechanism under a traditional LSTM prediction architecture, calculating the importance of critical trajectory points. Finally, our experiments conducted on the urban traffic dataset ApolloSpace validate the effectiveness of our proposed model. We demonstrate that MESTGN shows a significant performance improvement in vehicle trajectory prediction compared with existing mainstream models, thereby confirming its increased prediction accuracy.

Research on Efficient Asymmetric Attention Module for Real-Time Semantic Segmentation Networks in Urban Scenes

Currently, numerous high-precision models have been proposed for semantic segmentation, but the model parameters are large and the segmentation speed is slow. Real-time semantic segmentation for urban scenes necessitates a balance between accuracy, inference speed, and model size. In this paper, we present an efficient solution to this challenge, efficient asymmetric attention module net (EAAMNet) for the semantic segmentation of urban scenes, which adopts an asymmetric encoder–decoder structure. The encoder part of the network utilizes an efficient asymmetric attention module to form the network backbone. In the decoding part, we propose a lightweight multi-feature fusion decoder that can maintain good segmentation accuracy with a small number of parameters. Our extensive evaluations demonstrate that EAAMNet achieves a favorable equilibrium between segmentation efficiency, model parameters, and segmentation accuracy, rendering it highly suitable for real-time semantic segmentation in urban scenes. Remarkably, EAAMNet attains a 73.31% mIoU at 128 fps on Cityscapes and a 69.32% mIoU at 141 fps on CamVid without any pre-training. Compared to state-of-the-art models, our approach not only matches their model parameters but also enhances accuracy and increases speed.

CC-DETR: DETR with Hybrid Context and Multi-Scale Coordinate Convolution for Crowd Counting

Prevailing crowd counting approaches primarily rely on density map regression methods. Despite wonderful progress, significant scale variations and complex background interference within the same image remain challenges. To address these issues, in this paper we propose a novel DETR-based crowd counting framework called Crowd Counting DETR (CC-DETR), which aims to extend the state-of-the-art DETR object detection framework to the crowd counting task. In CC-DETR, a DETR-like encoder–decoder structure (Hybrid Context DETR, i.e., HCDETR) is proposed to tackle complex visual information by fusing features from hybrid semantic levels through a transformer. In addition, we design a Coordinate Dilated Convolution Module (CDCM) to effectively employ position-sensitive context information in different scales. Extensive experiments on three challenging crowd counting datasets (ShanghaiTech, UCF-QNRF, and NWPU) demonstrate that our model is effective and competitive when compared against SOTA crowd counting models.