Dong Zhou
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Dong Zhou has a total of 10 co-authors. Here are the top co-authors by number of shared publications:
- Wenjie Xie
Institute of Electronic Science and Technology, University of Electronic Science and Technology of China, Chengdu 611731, China
3 shared publications - Wenshuai Zhang
Institute of Electronic Science and Technology, University of Electronic Science and Technology of China, Chengdu 611731, China
3 shared publications - Dan Tian
Research Institute of Electronic Science and Technology, University of Electronic Science and Technology of China, Chengdu, China
2 shared publications - Wenrui Wang
The Research Institute of Electronic Science and Technology, University of Electronic Science and Technology of China, Chengdu 610054, China
2 shared publications - Chen Wang
University of Electronic Science and Technology of China, Chengdu, China
1 shared publication - Feng Guo
Research Institute of Electronic Science and Technology, University of Electronic Science and Technology of China, Chengdu 611731, China
1 shared publication
Latest Publications
- Open Access
- 3 June 2026
- Published by MDPI AG in Electronics
- Vol. 15 (11) , 2442
An Energy-Efficient FPGA-Based CNN Accelerator with Dual-Multiply Packing and Ping-Pong Buffering for Real-Time Object Detection
Real-time deployment of modern object-detection networks on edge devices is challenging because of limited compute resources, external-memory bandwidth, and strict power constraints. To address these issues, this paper presents a host–FPGA collaborative accelerator for quantized YOLOv5n on a Xilinx Zynq-7100 platform. The proposed design includes a modular multi-operator neural processing unit supporting seven atomic operators, a Dual-Multiply Packing (DMP) scheme to improve DSP48E1-based INT8 convolution density, a cache–compute–cache dataflow with global ping-pong buffering to overlap DMA transfers and computation, and a Multi-Quantization Domain Alignment (MQDA) pipeline to preserve accuracy at Add and Cat fusion nodes. Implemented at 200 MHz, the prototype achieves 24.617 ms FPGA-side forward-inference latency, 36.686 ms end-to-end single-frame latency, 27.2 FPS system-level performance, 182.8 GOPS equivalent throughput, and 8.536 W on-chip power consumption, corresponding to 21.42 GOPS/W. Experimental results also show that INT8 quantization causes only limited accuracy degradation, while MQDA improves quantized detection accuracy by reducing cross-domain fusion error. These results demonstrate that the proposed architecture provides an effective balance among throughput, energy efficiency, hardware cost, and quantized accuracy for real-time edge object detection.
- 3 June 2026
- Published by MDPI AG in Electronics
- Vol. 15 (11) , 2442
Open AccessAn Energy-Efficient FPGA-Based CNN Accelerator with Dual-Multiply Packing and Ping-Pong Buffering for Real-Time Object Detection
Real-time deployment of modern object-detection networks on edge devices is challenging because of limited compute resources, external-memory bandwidth, and strict power constraints. To address these issues, this paper presents a host–FPGA collaborative accelerator for quantized YOLOv5n on a Xilinx Zynq-7100 platform. The proposed design includes a modular multi-operator neural processing unit supporting seven atomic operators, a Dual-Multiply Packing (DMP) scheme to improve DSP48E1-based INT8 convolution density, a cache–compute–cache dataflow with global ping-pong buffering to overlap DMA transfers and computation, and a Multi-Quantization Domain Alignment (MQDA) pipeline to preserve accuracy at Add and Cat fusion nodes. Implemented at 200 MHz, the prototype achieves 24.617 ms FPGA-side forward-inference latency, 36.686 ms end-to-end single-frame latency, 27.2 FPS system-level performance, 182.8 GOPS equivalent throughput, and 8.536 W on-chip power consumption, corresponding to 21.42 GOPS/W. Experimental results also show that INT8 quantization causes only limited accuracy degradation, while MQDA improves quantized detection accuracy by reducing cross-domain fusion error. These results demonstrate that the proposed architecture provides an effective balance among throughput, energy efficiency, hardware cost, and quantized accuracy for real-time edge object detection.
- Open Access
SSRT-DETR: Domain-Adaptive Semi-Supervised Detector
Domain-adaptive object detection under set-prediction paradigms remains challenging, as Hungarian matching is sensitive to domain shift and fixed pseudo-label thresholds cannot simultaneously handle class imbalance and scene variability. This paper presents SSRT-DETR, a semi-supervised, domain-adaptive framework built on the real-time detector RT-DETR. We adopt a mean teacher–student architecture with style-transferred images to jointly model source and target domains. To stabilize the assignment process during the early stages of cross-domain training, Domain-Aware Matching (DAM) is formulated to augment the Hungarian matching cost with a teacher-guided decoder-query consistency term. Leveraging the more stable EMA teacher representations, DAM guides early matching toward domain-consistent assignments and is gradually annealed to recover standard matching as training converges. In parallel, we introduce Class-/Scene-Adaptive Pseudo-Labeling (CAP) to address a key limitation of existing DAOD methods that rely on fixed or globally tuned pseudo-label thresholds, which struggle with class imbalance and scene-dependent difficulty under domain shift. CAP leverages per-class confidence statistics and multi-view consistency to adapt classification and IoU thresholds across classes and scenes, while temperature scaling and quality-weighted losses provide soft control over pseudo-label reliability. Experiments on standard benchmarks demonstrate the robustness of SSRT-DETR. On Cityscapes→Foggy Cityscapes, SSRT-DETR improves mAP@0.5 from 51.0 to 54.3. On KITTI→Cityscapes and Sim10K→Cityscapes, it achieves 67.3 AP and 64.9 AP on the car category, respectively, clearly outperforming the RT-DETR baseline while maintaining real-time efficiency. Notably, consistent gains are observed in rare categories and adverse weather scenarios, validating the effectiveness of the proposed DAM and CAP modules.
Open AccessSSRT-DETR: Domain-Adaptive Semi-Supervised Detector
Domain-adaptive object detection under set-prediction paradigms remains challenging, as Hungarian matching is sensitive to domain shift and fixed pseudo-label thresholds cannot simultaneously handle class imbalance and scene variability. This paper presents SSRT-DETR, a semi-supervised, domain-adaptive framework built on the real-time detector RT-DETR. We adopt a mean teacher–student architecture with style-transferred images to jointly model source and target domains. To stabilize the assignment process during the early stages of cross-domain training, Domain-Aware Matching (DAM) is formulated to augment the Hungarian matching cost with a teacher-guided decoder-query consistency term. Leveraging the more stable EMA teacher representations, DAM guides early matching toward domain-consistent assignments and is gradually annealed to recover standard matching as training converges. In parallel, we introduce Class-/Scene-Adaptive Pseudo-Labeling (CAP) to address a key limitation of existing DAOD methods that rely on fixed or globally tuned pseudo-label thresholds, which struggle with class imbalance and scene-dependent difficulty under domain shift. CAP leverages per-class confidence statistics and multi-view consistency to adapt classification and IoU thresholds across classes and scenes, while temperature scaling and quality-weighted losses provide soft control over pseudo-label reliability. Experiments on standard benchmarks demonstrate the robustness of SSRT-DETR. On Cityscapes→Foggy Cityscapes, SSRT-DETR improves mAP@0.5 from 51.0 to 54.3. On KITTI→Cityscapes and Sim10K→Cityscapes, it achieves 67.3 AP and 64.9 AP on the car category, respectively, clearly outperforming the RT-DETR baseline while maintaining real-time efficiency. Notably, consistent gains are observed in rare categories and adverse weather scenarios, validating the effectiveness of the proposed DAM and CAP modules.
- Open Access
EAFormer: Edge-Aware Guided Adaptive Frequency-Navigator Network for Image Restoration
Although many deep learning-based image restoration networks have emerged in various image restoration tasks, most can only perform well in a specific type of restoration task and still face challenges in the general performance of image restoration. The fundamental reason for this problem is that different types of degradation require different frequency features, and the image needs to be adaptively reconstructed according to the characteristics of input degradation. At the same time, we noticed that the previous image restoration network ignored the reconstruction of the edge contour details of the image, resulting in unclear contours of the restored image. Therefore, we proposed an edge-aware guided adaptive frequency navigation network, EAFormer, which extracts edge information in the image by applying different edge detection operators and reconstructs the edge contour details of the image more accurately during the restoration process. The adaptive frequency navigation perceives different frequency components in the image and interactively participates in the subsequent restoration process with high- and low-frequency feature information, better retaining the global structural information of the image and making the restored image more visually coherent and realistic. We verified the versatility of EAFormer in five classic image restoration tasks, and many experimental results also show that our model has advanced performance.
Open AccessEAFormer: Edge-Aware Guided Adaptive Frequency-Navigator Network for Image Restoration
Although many deep learning-based image restoration networks have emerged in various image restoration tasks, most can only perform well in a specific type of restoration task and still face challenges in the general performance of image restoration. The fundamental reason for this problem is that different types of degradation require different frequency features, and the image needs to be adaptively reconstructed according to the characteristics of input degradation. At the same time, we noticed that the previous image restoration network ignored the reconstruction of the edge contour details of the image, resulting in unclear contours of the restored image. Therefore, we proposed an edge-aware guided adaptive frequency navigation network, EAFormer, which extracts edge information in the image by applying different edge detection operators and reconstructs the edge contour details of the image more accurately during the restoration process. The adaptive frequency navigation perceives different frequency components in the image and interactively participates in the subsequent restoration process with high- and low-frequency feature information, better retaining the global structural information of the image and making the restored image more visually coherent and realistic. We verified the versatility of EAFormer in five classic image restoration tasks, and many experimental results also show that our model has advanced performance.
- Open Access
- 26 March 2025
- Published by MDPI AG in Applied Sciences
- Vol. 15 (7) , 3627
Few-Shot Semantic Segmentation Network for Distinguishing Positive and Negative Examples
Few-shot segmentation (FSS) aims to segment a query image with a few support images. However, there can be large differences between images from the same category, and similarities between different categories, making it a challenging task. In addition, most FSS methods use powerful encoders to extract features from the training class, which makes the model pay more attention to the features of the ‘seen’ class, and perform poorly on the segmentation task of ‘unseen’ classes. In this work, we propose a novel end-to-end model, called GFormer. GFormer has four components: encoder, prototype extractor, adversarial prototype generator, and decoder. Our encoder makes simple modifications to VIT to reduce the focus on image content, using a prototype extractor to extract prototype features from a range of support images. We further introduce different classes that are similar to the support image categories as negative examples, taking the support image categories as positive examples. We use the adversarial prototype generator to extract the adversarial prototypes from the positive and negative examples. The decoder segments the query images under the guidance of the prototypes. We conduct extensive experiments on a variety of unknown classes. The results verify the feasibility of the proposed model and prove that the proposed model has strong generalization performance for new classes.
- 26 March 2025
- Published by MDPI AG in Applied Sciences
- Vol. 15 (7) , 3627
Open AccessFew-Shot Semantic Segmentation Network for Distinguishing Positive and Negative Examples
Few-shot segmentation (FSS) aims to segment a query image with a few support images. However, there can be large differences between images from the same category, and similarities between different categories, making it a challenging task. In addition, most FSS methods use powerful encoders to extract features from the training class, which makes the model pay more attention to the features of the ‘seen’ class, and perform poorly on the segmentation task of ‘unseen’ classes. In this work, we propose a novel end-to-end model, called GFormer. GFormer has four components: encoder, prototype extractor, adversarial prototype generator, and decoder. Our encoder makes simple modifications to VIT to reduce the focus on image content, using a prototype extractor to extract prototype features from a range of support images. We further introduce different classes that are similar to the support image categories as negative examples, taking the support image categories as positive examples. We use the adversarial prototype generator to extract the adversarial prototypes from the positive and negative examples. The decoder segments the query images under the guidance of the prototypes. We conduct extensive experiments on a variety of unknown classes. The results verify the feasibility of the proposed model and prove that the proposed model has strong generalization performance for new classes.
- 20 February 2025
- Published by Springer Nature in The Visual Computer
- Vol. 41 (10) , 7799-7815
Multi-scale feature fusion network for real-time semantic segmentation of urban street scenes: enhancing detail retention and accuracy
- 20 February 2025
- Published by Springer Nature in The Visual Computer
- Vol. 41 (10) , 7799-7815
Multi-scale feature fusion network for real-time semantic segmentation of urban street scenes: enhancing detail retention and accuracy
- Open Access
IV-YOLO: A Lightweight Dual-Branch Object Detection Network
With the rapid growth in demand for security surveillance, assisted driving, and remote sensing, object detection networks with robust environmental perception and high detection accuracy have become a research focus. However, single-modality image detection technologies face limitations in environmental adaptability, often affected by factors such as lighting conditions, fog, rain, and obstacles like vegetation, leading to information loss and reduced detection accuracy. We propose an object detection network that integrates features from visible light and infrared images—IV-YOLO—to address these challenges. This network is based on YOLOv8 (You Only Look Once v8) and employs a dual-branch fusion structure that leverages the complementary features of infrared and visible light images for target detection. We designed a Bidirectional Pyramid Feature Fusion structure (Bi-Fusion) to effectively integrate multimodal features, reducing errors from feature redundancy and extracting fine-grained features for small object detection. Additionally, we developed a Shuffle-SPP structure that combines channel and spatial attention to enhance the focus on deep features and extract richer information through upsampling. Regarding model optimization, we designed a loss function tailored for multi-scale object detection, accelerating the convergence speed of the network during training. Compared with the current state-of-the-art Dual-YOLO model, IV-YOLO achieves mAP improvements of 2.8%, 1.1%, and 2.2% on the Drone Vehicle, FLIR, and KAIST datasets, respectively. On the Drone Vehicle and FLIR datasets, IV-YOLO has a parameter count of 4.31 M and achieves a frame rate of 203.2 fps, significantly outperforming YOLOv8n (5.92 M parameters, 188.6 fps on the Drone Vehicle dataset) and YOLO-FIR (7.1 M parameters, 83.3 fps on the FLIR dataset), which had previously achieved the best performance on these datasets. This demonstrates that IV-YOLO achieves higher real-time detection performance while maintaining lower parameter complexity, making it highly promising for applications in autonomous driving, public safety, and beyond.
Open AccessIV-YOLO: A Lightweight Dual-Branch Object Detection Network
With the rapid growth in demand for security surveillance, assisted driving, and remote sensing, object detection networks with robust environmental perception and high detection accuracy have become a research focus. However, single-modality image detection technologies face limitations in environmental adaptability, often affected by factors such as lighting conditions, fog, rain, and obstacles like vegetation, leading to information loss and reduced detection accuracy. We propose an object detection network that integrates features from visible light and infrared images—IV-YOLO—to address these challenges. This network is based on YOLOv8 (You Only Look Once v8) and employs a dual-branch fusion structure that leverages the complementary features of infrared and visible light images for target detection. We designed a Bidirectional Pyramid Feature Fusion structure (Bi-Fusion) to effectively integrate multimodal features, reducing errors from feature redundancy and extracting fine-grained features for small object detection. Additionally, we developed a Shuffle-SPP structure that combines channel and spatial attention to enhance the focus on deep features and extract richer information through upsampling. Regarding model optimization, we designed a loss function tailored for multi-scale object detection, accelerating the convergence speed of the network during training. Compared with the current state-of-the-art Dual-YOLO model, IV-YOLO achieves mAP improvements of 2.8%, 1.1%, and 2.2% on the Drone Vehicle, FLIR, and KAIST datasets, respectively. On the Drone Vehicle and FLIR datasets, IV-YOLO has a parameter count of 4.31 M and achieves a frame rate of 203.2 fps, significantly outperforming YOLOv8n (5.92 M parameters, 188.6 fps on the Drone Vehicle dataset) and YOLO-FIR (7.1 M parameters, 83.3 fps on the FLIR dataset), which had previously achieved the best performance on these datasets. This demonstrates that IV-YOLO achieves higher real-time detection performance while maintaining lower parameter complexity, making it highly promising for applications in autonomous driving, public safety, and beyond.