YOLOv7-based small object detection model
Zhu Fan
Xihua University
DOI: https://doi.org/10.59429/esta.v12i3.11366
Keywords: object detection; YOLOv7; knowledge distillation; lightweight model; multi-scale feature transfer
Abstract
Aiming at the problems of high model computation cost and easy feature loss during lightweighting in small target detection task, this paper proposes a compact feature migration method based on YOLOv7, which reduces the number of model parameters by combining with knowledge distillation technique and can maintain relatively good detection effect. The method reduces the number of model parameters by designing a lightweight student network through channel pruning; then dynamically assigns spatial weights to enhance the feature alignment in small target regions through the Adaptive Feature Distillation (AFD) module; and finally, further compresses the model by combining structured pruning with quantized perceptual training. The experimental results show that this paper's method decreases the model mAP by only 1.8% with 35% reduction in the amount of parameters and 53% reduction in the amount of computation. It can be seen that the method provides an effective lightweight solution for small target detection, which has certain practical significance.
References
[1] Zeni L F, Jung C R. Distilling knowledge from refinement in multiple instance detection networks[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. 2020: 768-769.
[2] Farhadi M, Ghasemi M, Vrudhula S, et al. Enabling incremental knowledge transfer for object detection at the edge[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. 2020: 396-397.
[3] Zheng Z, Ye R, Hou Q, et al. Localization distillation for object detection[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023, 45(8): 10070-10083.
[4] Dai X, Jiang Z, Wu Z, et al. General instance distillation for object detection[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2021: 7842-7851.
[5] Wang C Y, Bochkovskiy A, Liao H Y M. YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2023: 7464-7475.