Hybrid Architecture Edge Object Detector Based on Structured Pruning and Attention Reparameterization
DOI:
https://doi.org/10.67541/jdc2604Keywords:
Edge object detection; Structured pruning; Attention reparameterization; Hybrid architecture; Model compressionAbstract
Edge computing scenarios put forward strict requirements on inference delay and computational power consumption of target detectors, and the existing lightweight models are difficult to achieve the best balance between accuracy and speed. In this paper, a hybrid architecture edge object detector based on structured pruning and attention reparameterization is proposed, which is co-optimized from three levels: architecture design, model compression and inference acceleration. Firstly, a CNN-Transformer dual-branch heterogeneous backbone network is constructed, and the basic mAP is increased to 44.2% on the premise of only increasing the number of parameters by 15% by fusing local details and global context features through cross-modal interaction modules. Then, a structured pruning algorithm driven by hierarchical sensitivity is proposed, which combines Hessian trace and batch normalization scaling factor to dynamically determine the redundancy of each layer. The differentiable mask and Group Lasso sparse regularity are introduced to realize training, namely pruning. With the layer-by-layer feature reconstruction compensation mechanism, the accuracy retention rate can reach 97.5% when FLOPs are reduced by 50.4%. The attention reparameterization technique during inference is further proposed, which uses the linear additivity of affine transformations to losslessly fuse the multi-branch attention structure in the training stage into a single convolutional layer, and combines piecewise linear fitting to deal with Softmax nonlinearity, completely eliminating the attention computation cost during inference. Experiments show that the proposed method achieves 43.0% mAP on COCO dataset, 112.3 FPS on Jetson Xavier, and the model size is only 5.8 MB. In terms of accuracy, speed and model size, it is superior to mainstream methods such as YOLOv8n, MobileDet and EfficientDet-D0, and also shows optimal robustness in edge scenes such as low light and motion blur.
References
[1] Khanam, R., Hussain, M., Hill, R., & Allen, P. (2024). A comprehensive review of convolutional neural networks for defect detection in industrial applications. Ieee Access, 12, 94250-94295. https://doi.org/10.1109/ACCESS.2024.3425166
[2] Elvander, S. (2026). Deep Learning-Enabled Industrial Intelligence for Smart Inspection, Equipment Diagnosis, Predictive Maintenance, and Edge Deployment. Journal of Computer Technology and Software, 5(4). https://doi.org/10.1109/ICICT60155.2024.10544377
[3] Sreejith, R., Gupta, T. R., Yamsani, N., Singh, A., Haldorai, A., & Hussain, T. (2026, February). LightClassNet-A Lightweight Convolutional Model for Efficient Image Classification on Edge Devices. In 2026 International Conference on ICT and Photonics (ICTP) (Vol. 1, pp. 1-6). IEEE. https://doi.org/10.1109/ICTP67998.2026.11485470
[4] Appavu, N. (2025, August). Analysing the effect of edge-optimized deep learning models on improving low-powered iot devices real-time object detection. In 2025 9th International Conference on Inventive Systems and Control (ICISC) (pp. 1663-1669). IEEE. https://doi.org/10.1109/ICISC65841.2025.11188323
[5] Lin, B., Wang, J., Wang, H., Zhong, L., Yang, X., & Zhang, X. (2023). Small space target detection based on a convolutional neural network and guidance information. Aerospace, 10(5), 426. https://doi.org/10.3390/aerospace10050426
[6] Su, S., Niu, W., Li, Y., Ren, C., Peng, X., Zheng, W., & Yang, Z. (2023). Dim and small space-target detection and centroid positioning based on motion feature learning. Remote Sensing, 15(9), 2455. https://doi.org/10.3390/rs15092455
[7] Chen, B., Liu, L., Zou, Z., & Shi, Z. (2023). Target detection in hyperspectral remote sensing image: Current status and challenges. Remote Sensing, 15(13), 3223. https://doi.org/10.3390/rs15133223
[8] Luan, D., & Thompson, J. S. (2023). Channelformer: Attention based neural solution for wireless channel estimation and effective online training. IEEE Transactions on Wireless Communications, 22(10), 6562-6577. https://doi.org/10.1109/TWC.2023.3244484
[9] Hu, J., Liu, Y., & Wu, K. (2022). Neural network pruning based on channel attention mechanism. Connection Science, 34(1), 2201-2218. https://doi.org/10.1109/CECIT58139.2022.00042
[10] Lian, D., Zhou, D., Feng, J., & Wang, X. (2022). Scaling & shifting your features: A new baseline for efficient model tuning. Advances in Neural Information Processing Systems, 35, 109-123. https://doi.org/10.48550/arXiv.2210.08823
[11] Ding, N., Qin, Y., Yang, G., Wei, F., Yang, Z., Su, Y., ... & Sun, M. (2023). Parameter-efficient fine-tuning of large-scale pre-trained language models. Nature machine intelligence, 5(3), 220-235. https://doi.org/10.1038/s42256-023-00626-4
[12] Fan, Y., Xu, W., Fushuo, H., Cheng, N., Haozhao, W., Zhou, H., ... & Shen, X. (2025). Inter-modal Interactions in Multimodal Learning: A Comprehensive Survey. ACM Comput Surv, 37(4). https://doi.org/10.48550/arXiv.2411.17040
[13] Sharma, P., & Jain, S. K. (2026). A survey of transformers based on their input modalities. Applied Intelligence, 56(10), 348. https://doi.org/10.1007/s10489-026-07378-9
[14] Hou, L., Zhuang, Y., Xie, Y., Kan, H., Huang, Z., & Lin, J. (2025). Cross-modal generalizable visual-language models via inter-modal bidirectional supervision for enhanced pathology image recognition. Pattern Recognition, 112240. https://doi.org/10.1016/j.patcog.2025.112240
[15] Song, L., Xia, M., Weng, L., Lin, H., Qian, M., & Chen, B. (2022). Axial cross attention meets CNN: Bibranch fusion network for change detection. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 16, 21-32. https://doi.org/10.1109/JSTARS.2022.3224081
[16] Badhe, N. B., Yele, V. P., Neve, R. P., & Tinsu, J. (2025). Hybrid three branch CNN and transformer with skeleton based attentions network for change detection in remote sensing images. Sensing and Imaging, 26(1), 109. https://doi.org/10.1007/s11220-025-00632-3
[17] Cao, M., Wang, L., Zhu, M., & Yuan, X. (2024). Hybrid CNN-transformer architecture for efficient large-scale video snapshot compressive imaging. International Journal of Computer Vision, 132(10), 4521-4540. https://doi.org/10.1007/s11263-024-02101-y
[18] Wang, Q., Yin, C., She, K., Tong, Q., Lu, G., Zhang, H., & Lu, J. (2025). Bearing fault diagnosis for variable operating conditions based on KAN convolution and dual branch fusion attention. Scientific Reports, 15(1), 21442. https://doi.org/10.1038/s41598-025-04620-1
[19] Song, X., Cong, Y., Song, Y., Chen, Y., & Liang, P. (2022). A bearing fault diagnosis model based on CNN with wide convolution kernels: X. Song et al. Journal of Ambient Intelligence and Humanized Computing, 13(8), 4041-4056. https://doi.org/10.1007/s12652-021-03177-x
[20] Huang, Q., & Huang, J. (2025). Comprehensive review of edge and contour detection: From traditional methods to recent advances. Neural Computing and Applications, 37(4), 2175-2209. https://doi.org/10.1007/s00521-024-10936-2
[21] Ghojogh, B., & Ghodsi, A. (2026). Attention mechanism and transformers. In Elements of deep learning (pp. 231-257). Cham: Springer Nature Switzerland. https://doi.org/10.1007/978-3-032-10738-1_9
[22] Kohan, A., Rietman, E. A., & Siegelmann, H. T. (2023). Signal propagation: The framework for learning and inference in a forward pass. IEEE Transactions on Neural Networks and Learning Systems, 35(6), 8585-8596. https://doi.org/10.1109/TNNLS.2022.3230914
[23] Xiaosen, W., Tong, K., & He, K. (2023). Rethinking the backward propagation for adversarial transferability. Advances in Neural Information Processing Systems, 36, 1905-1922. https://doi.org/10.48550/arXiv.2306.12685
[24] Shi, J., Wang, Y., Yu, Z., Li, G., Hong, X., Wang, F., & Gong, Y. (2023). Exploiting multi-scale parallel self-attention and local variation via dual-branch transformer-CNN structure for face super-resolution. IEEE Transactions on Multimedia, 26, 2608-2620. https://doi.org/10.1109/TMM.2023.3301225
[25] Zhang, H., Du, Q., Qi, Q., Zhang, J., Wang, F., & Gao, M. (2023). A recursive attention-enhanced bidirectional feature pyramid network for small object detection. Multimedia tools and applications, 82(9), 13999-14018. https://doi.org/10.1007/s11042-022-13951-4
[26] Tao, S., Fu, H., Yang, R., & Wang, L. (2025). A multi-task learning framework with enhanced cross-level semantic consistency for multi-level land cover classification. Remote Sensing, 17(14), 2442. https://doi.org/10.3390/rs17142442
[27] Liu, L., Pan, Y., & Tu, B. (2026). BDSANet: Dual-branch combined directional sparse attention mechanism model for road network extraction. IEEE Transactions on Geoscience and Remote Sensing, 64, 1-16. https://doi.org/10.1109/TGRS.2026.3678070
[28] Liu, T., Luo, R., Xu, L., Feng, D., Cao, L., Liu, S., & Guo, J. (2022). Spatial channel attention for deep convolutional neural networks. Mathematics, 10(10), 1750. https://doi.org/10.3390/math10101750
Published
Data Availability Statement
The data that support the findings of this study are available upon request from the corresponding authors, R.Z.
Issue
Section
License
Copyright (c) 2026 Ruisheng Zhang (Author)
This work is licensed under a Creative Commons Attribution 4.0 International License.
