Junzhu Mao
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Junzhu Mao has a total of 21 co-authors. Here are the top co-authors by number of shared publications:
- Yazhou Yao
Nanjing University of Science and Technology, Nanjing, China
8 shared publications - Zeren Sun
Nanjing University of Science and Technology
4 shared publications - Heng Tao Shen
Tongji University, Shanghai, China
3 shared publications - Xian-Sheng Hua
Tongji University, Shanghai, China
3 shared publications - Fumin Shen
School of Computer Science and Engineering, University of Electronic Science and Technology of China, Chengdu, China
2 shared publications - Gensheng Pei
Sungkyunkwan University
2 shared publications
Latest Publications
- 12 September 2026
- Published by Springer Nature
- p. 469-490
Beyond Attention: Convolutional Global Context for Remote Sensing Change Detection
- 12 September 2026
- Published by Springer Nature
- p. 469-490
Beyond Attention: Convolutional Global Context for Remote Sensing Change Detection
- 12 September 2026
- Published by Springer Nature
- p. 19-36
From Reasoning to Pixels: Grounded Medical Multimodal LLMs for VQA and Segmentation
- 12 September 2026
- Published by Springer Nature
- p. 19-36
From Reasoning to Pixels: Grounded Medical Multimodal LLMs for VQA and Segmentation
- 19 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Multimedia
- Vol. PP (15209210) , 1-12
Class Importance Consistency Matters: Efficient Model Pruning for Long-tailed Recognition Models
Model pruning, aiming to reduce computational costs for deep neural networks, is essential for deploying deep models in resource-limited scenarios. While significant progress has been witnessed in pruning techniques in recent years, existing methods predominantly focus on balanced datasets. Adapting these approaches to datasets with long-tailed data distributions remains challenging since their pronounced class imbalances often lead to under-learning of tail-class samples, resulting in inaccurate estimation for parameter importance. This causes notable performance degradation after pruning. To this end, we propose Class Importance Consistency (CIC) pruning by introducing a consistency-guided criterion that enhances parameter importance evaluation across all classes. By prioritizing model parameters with stable importance across all classes and pruning those with inconsistent contributions, CIC effectively mitigates the effects of skewed importance scores. Additionally, we introduce a Class-wise Batch Sampling (CBS) strategy to streamline the computation of class-wise importance scores, significantly improving pruning efficiency. Extensive experiments on CIFAR100-LT, ImageNet-LT, and iNaturalist2018 demonstrate that CIC consistently achieves Pareto-optimal parameter-accuracy trade-offs, surpassing state-of-the-art methods in average accuracy by a margin over 1%.
- 19 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Multimedia
- Vol. PP (15209210) , 1-12
Class Importance Consistency Matters: Efficient Model Pruning for Long-tailed Recognition Models
Model pruning, aiming to reduce computational costs for deep neural networks, is essential for deploying deep models in resource-limited scenarios. While significant progress has been witnessed in pruning techniques in recent years, existing methods predominantly focus on balanced datasets. Adapting these approaches to datasets with long-tailed data distributions remains challenging since their pronounced class imbalances often lead to under-learning of tail-class samples, resulting in inaccurate estimation for parameter importance. This causes notable performance degradation after pruning. To this end, we propose Class Importance Consistency (CIC) pruning by introducing a consistency-guided criterion that enhances parameter importance evaluation across all classes. By prioritizing model parameters with stable importance across all classes and pruning those with inconsistent contributions, CIC effectively mitigates the effects of skewed importance scores. Additionally, we introduce a Class-wise Batch Sampling (CBS) strategy to streamline the computation of class-wise importance scores, significantly improving pruning efficiency. Extensive experiments on CIFAR100-LT, ImageNet-LT, and iNaturalist2018 demonstrate that CIC consistently achieves Pareto-optimal parameter-accuracy trade-offs, surpassing state-of-the-art methods in average accuracy by a margin over 1%.
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Multimedia
- Vol. 28, 15866-15876
Selective Timestep Pruning for Efficient Diffusion Models
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Multimedia
- Vol. 28, 15866-15876
Selective Timestep Pruning for Efficient Diffusion Models
- 17 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Image Processing
- Vol. 34 (10577149) , 8447-8460
NiCI-Pruning: Enhancing Diffusion Model Pruning via Noise in Clean Image Guidance
The substantial successes achieved by diffusion probabilistic models have prompted the study of their employment in resource-limited scenarios. Pruning methods have been proven effective in compressing discriminative models relying on the correlation between training losses and model performances. However, diffusion models employ an iterative process for generating high-quality images, leading to a breakdown of such connections. To address this challenge, we propose a simple yet effective method, named NiCI-Pruning (Noise in Clean Image Pruning), for the compression of diffusion models. NiCI-Pruning capitalizes the noise predicted by the model based on clean image inputs, favoring it as a feature for establishing reconstruction losses. Accordingly, Taylor expansion is employed for the proposed reconstruction loss to evaluate the parameter importance effectively. Moreover, we propose an interval sampling strategy that incorporates a timestep-weighted schema, alleviating the risk of misleading information obtained at later timesteps. We provide comprehensive experimental results to affirm the superiority of our proposed approach. Notably, our method achieves a remarkable average reduction of 30.4% in FID score increase across five different datasets compared to the state-of-the-art diffusion pruning method at equivalent pruning rates. Our code and models have been made available at https://github.com/ NUST-Machine-Intelligence-Laboratory/NiCI-Pruning.
- 17 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Image Processing
- Vol. 34 (10577149) , 8447-8460
NiCI-Pruning: Enhancing Diffusion Model Pruning via Noise in Clean Image Guidance
The substantial successes achieved by diffusion probabilistic models have prompted the study of their employment in resource-limited scenarios. Pruning methods have been proven effective in compressing discriminative models relying on the correlation between training losses and model performances. However, diffusion models employ an iterative process for generating high-quality images, leading to a breakdown of such connections. To address this challenge, we propose a simple yet effective method, named NiCI-Pruning (Noise in Clean Image Pruning), for the compression of diffusion models. NiCI-Pruning capitalizes the noise predicted by the model based on clean image inputs, favoring it as a feature for establishing reconstruction losses. Accordingly, Taylor expansion is employed for the proposed reconstruction loss to evaluate the parameter importance effectively. Moreover, we propose an interval sampling strategy that incorporates a timestep-weighted schema, alleviating the risk of misleading information obtained at later timesteps. We provide comprehensive experimental results to affirm the superiority of our proposed approach. Notably, our method achieves a remarkable average reduction of 30.4% in FID score increase across five different datasets compared to the state-of-the-art diffusion pruning method at equivalent pruning rates. Our code and models have been made available at https://github.com/ NUST-Machine-Intelligence-Laboratory/NiCI-Pruning.
- 30 July 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Consumer Electronics
- Vol. 71 (3) , 8880-8891
Advancing Industrial Battery Defect Analysis via Multi-Domain Feature Integration and Optimized Neural Pruning for Edge System
Automated battery defect detection is essential for ensuring quality and safety in consumer electronics manufacturing, enabling enhanced edge-level decision-making and operational efficiency. However, existing methods struggle to achieve both high accuracy and real-time performance required by edge-based industrial scenarios. This paper introduces a lightweight, multi-domain feature integration approach designed specifically for edge deployment, significantly enhancing the practical applicability of battery defect classification. Firstly, a bilateral scanning projection module is developed for precise and rapid battery localization, segmenting battery images efficiently into inner circle, inner ring, and outer ring regions. Subsequently, a spatial-statistical feature fusion module effectively captures structural defects in the spatial domain. To streamline feature extraction and ensure consistent classification accuracy, the annular regions undergo polar coordinate transformation, resulting in simplified geometric representation. Additionally, we propose a frequency-enhanced optimization strategy tailored for lightweight edge networks, utilizing Fast Fourier Transform and frequency filtering methods to accurately identify complex texture defects. To further meet real-time requirements and minimize computational overhead, we apply Taylor pruning techniques to optimize network performance, facilitating efficient deployment using the OpenVINO framework on edge devices. The proposed lightweight AI framework achieves a classification accuracy of 98.8% while maintaining a rapid inference speed of 20ms per image, substantially outperforming current battery defect detection approaches and demonstrating significant potential for edge-enhanced decision-making in consumer electronics manufacturing. Our code is available at https://github.com/NUST-Machine-Intelligence-Laboratory/MDFI.
- 30 July 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Consumer Electronics
- Vol. 71 (3) , 8880-8891
Advancing Industrial Battery Defect Analysis via Multi-Domain Feature Integration and Optimized Neural Pruning for Edge System
Automated battery defect detection is essential for ensuring quality and safety in consumer electronics manufacturing, enabling enhanced edge-level decision-making and operational efficiency. However, existing methods struggle to achieve both high accuracy and real-time performance required by edge-based industrial scenarios. This paper introduces a lightweight, multi-domain feature integration approach designed specifically for edge deployment, significantly enhancing the practical applicability of battery defect classification. Firstly, a bilateral scanning projection module is developed for precise and rapid battery localization, segmenting battery images efficiently into inner circle, inner ring, and outer ring regions. Subsequently, a spatial-statistical feature fusion module effectively captures structural defects in the spatial domain. To streamline feature extraction and ensure consistent classification accuracy, the annular regions undergo polar coordinate transformation, resulting in simplified geometric representation. Additionally, we propose a frequency-enhanced optimization strategy tailored for lightweight edge networks, utilizing Fast Fourier Transform and frequency filtering methods to accurately identify complex texture defects. To further meet real-time requirements and minimize computational overhead, we apply Taylor pruning techniques to optimize network performance, facilitating efficient deployment using the OpenVINO framework on edge devices. The proposed lightweight AI framework achieves a classification accuracy of 98.8% while maintaining a rapid inference speed of 20ms per image, substantially outperforming current battery defect detection approaches and demonstrating significant potential for edge-enhanced decision-making in consumer electronics manufacturing. Our code is available at https://github.com/NUST-Machine-Intelligence-Laboratory/MDFI.
- 27 January 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Multimedia
- Vol. 27 (15209210) , 4670-4683
Prune and Merge: Efficient Token Compression for Vision Transformer With Spatial Information Preserved
Token compression is essential for reducing the computational and memory requirements of transformer models, enabling their deployment in resource-constrained environments. In this work, we propose an efficient and hardware-compatible token compression method called Prune and Merge. Our approach integrates token pruning and merging operations within transformer models to achieve layer-wise token compression. By introducing trainable merge and reconstruct matrices and utilizing shortcut connections, we efficiently merge tokens while preserving important information and enabling the restoration of pruned tokens. Additionally, we introduce a novel gradient-weighted attention scoring mechanism that computes token importance scores during the training phase, eliminating the need for separate computations during inference and enhancing compression efficiency. We also leverage gradient information to capture the global impact of tokens and automatically identify optimal compression structures. Extensive experiments on the ImageNet-1k and ADE20K datasets validate the effectiveness of our approach, achieving significant speed-ups with minimal accuracy degradation compared to state-of-the-art methods. For instance, on DeiT-Small, we achieve a 1.64× speed-up with only a 0.2% drop in accuracy on ImageNet-1k. Moreover, by compressing segmenter models and comparing with existing methods, we demonstrate the superior performance of our approach in terms of efficiency and effectiveness. Code and models have been made available at https://github.com/NUST-Machine-Intelligence-Laboratory/prune_and_merge.
- 27 January 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Multimedia
- Vol. 27 (15209210) , 4670-4683
Prune and Merge: Efficient Token Compression for Vision Transformer With Spatial Information Preserved
Token compression is essential for reducing the computational and memory requirements of transformer models, enabling their deployment in resource-constrained environments. In this work, we propose an efficient and hardware-compatible token compression method called Prune and Merge. Our approach integrates token pruning and merging operations within transformer models to achieve layer-wise token compression. By introducing trainable merge and reconstruct matrices and utilizing shortcut connections, we efficiently merge tokens while preserving important information and enabling the restoration of pruned tokens. Additionally, we introduce a novel gradient-weighted attention scoring mechanism that computes token importance scores during the training phase, eliminating the need for separate computations during inference and enhancing compression efficiency. We also leverage gradient information to capture the global impact of tokens and automatically identify optimal compression structures. Extensive experiments on the ImageNet-1k and ADE20K datasets validate the effectiveness of our approach, achieving significant speed-ups with minimal accuracy degradation compared to state-of-the-art methods. For instance, on DeiT-Small, we achieve a 1.64× speed-up with only a 0.2% drop in accuracy on ImageNet-1k. Moreover, by compressing segmenter models and comparing with existing methods, we demonstrate the superior performance of our approach in terms of efficiency and effectiveness. Code and models have been made available at https://github.com/NUST-Machine-Intelligence-Laboratory/prune_and_merge.
- 6 April 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Multimedia
- Vol. 25 (15209210) , 1592-1599
Attention Map Guided Transformer Pruning for Occluded Person Re-Identification on Edge Device
Due to its significant capability of modeling long-range dependencies, vision transformer (ViT) has achieved promising success in both holistic and occluded person re-identification (Re-ID) tasks. However, the inherent problems of transformers such as the huge computational cost and memory footprint are still two unsolved issues that will block the deployment of ViT based person Re-ID models on resource-limited edge devices. Our goal is to reduce both the inference complexity and model size without sacrificing the comparable accuracy on person Re-ID, especially for tasks with occlusion. To this end, we propose a novel attention map guided (AMG) transformer pruning method, which removes both redundant tokens and heads with the guidance of the attention map in a hardware-friendly way. We first calculate the entropy in the key dimension and sum it up for the whole map, and the corresponding head parameters of maps with high entropy will be removed for model size reduction. Then we combine the similarity and first-order gradients of key tokens along the query dimension for token importance estimation and remove redundant key and value tokens to further reduce the inference complexity. Comprehensive experiments on Occluded DukeMTMC and Market-1501 demonstrate the effectiveness of our proposals. For example, our proposed pruning strategy on ViT-Base enjoys 29.4% FLOPs savings with 0.2% drop on Rank-1 and 0.4% improvement on mAP, respectively. Code and models have been made available at https://github.com/NUST-Machine-Intelligence-Laboratory/AMG .
- 6 April 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Multimedia
- Vol. 25 (15209210) , 1592-1599
Attention Map Guided Transformer Pruning for Occluded Person Re-Identification on Edge Device
Due to its significant capability of modeling long-range dependencies, vision transformer (ViT) has achieved promising success in both holistic and occluded person re-identification (Re-ID) tasks. However, the inherent problems of transformers such as the huge computational cost and memory footprint are still two unsolved issues that will block the deployment of ViT based person Re-ID models on resource-limited edge devices. Our goal is to reduce both the inference complexity and model size without sacrificing the comparable accuracy on person Re-ID, especially for tasks with occlusion. To this end, we propose a novel attention map guided (AMG) transformer pruning method, which removes both redundant tokens and heads with the guidance of the attention map in a hardware-friendly way. We first calculate the entropy in the key dimension and sum it up for the whole map, and the corresponding head parameters of maps with high entropy will be removed for model size reduction. Then we combine the similarity and first-order gradients of key tokens along the query dimension for token importance estimation and remove redundant key and value tokens to further reduce the inference complexity. Comprehensive experiments on Occluded DukeMTMC and Market-1501 demonstrate the effectiveness of our proposals. For example, our proposed pruning strategy on ViT-Base enjoys 29.4% FLOPs savings with 0.2% drop on Rank-1 and 0.4% improvement on mAP, respectively. Code and models have been made available at https://github.com/NUST-Machine-Intelligence-Laboratory/AMG .