| 摘要: |
| 目的 野火预警大多采用烟雾或红外传感器检测,且这些传感器在大型开放式空间下,容易受到环境的影
响,从而很难进行开放场所的精准火灾预警,而优越的火焰检测模型往往存在过多的参数量,且存在结构冗余的问
题,基于此问题,提出一种改进的 VGG 深度卷积网络架构。 方法 以映射变换为基础,进行像素值调整,在保证分类
精度的前提下,采用 L1 正则化保证稀疏性,并基于 BN 层进行结构化剪枝,从而降低模型储存数据量,得到精简的
模型。 结果 大量的仿真试验结果表明:该方法在不同剪枝比例下,在野火架构数据集上,检测与勘误率依然能够保
持高的准确精度,改进的模型在剪枝率为 80%时,准确率达到了 95. 29%,提升了 0. 92%,并有效解决了模型过参数
化的问题;通过不同的微调训练,模型精度略微超过没有进行剪枝时的模型,且在参数量上少了近 20 倍,并随着剪
枝率的上升,检测效果在原有精度水平上无明显下降,甚至略高于原始模型精度,这说明在训练过程中,有大量的
冗余权重。 结论 该方法可以大幅度缩减模型的储存量,并可保证较高的分类精度,具有较好的实际应用意义,可以
应用在神经网络存储计算能力较弱的嵌入式设备中。 |
| 关键词: 深度卷积网络 VGG 火焰检测 剪枝 |
| DOI： |
| 分类号: |
| 基金项目: |
|
| Fire Image Recognition Method Based on Sparse Convolutional Network Pruning |
|
YAN Jiawen LIN Xiankun PAN Yizhou
|
|
School of Mechanical Engineering University of Shanghai for Science and Technology Shanghai 200093 China
|
| Abstract: |
| Objective Under most circumstances wildfire warnings primarily rely on smoke or infrared sensors for
detection. However these sensors are susceptible to environmental interference especially in large open spaces making
it challenging to achieve precise fire alerts in open areas. Additionally superior flame detection models often have too
many parameters and suffer from structural redundancy. Based on the above problems an improved VGG deep
convolutional network architecture was proposed. Methods Pixel value adjustments were made based on mapping
transformations. While ensuring classification accuracy L1 regularization was employed to ensure sparsity. Structural
pruning was performed based on BN layers thereby reducing the storage data volume of the model and obtaining a
streamlined model. Results Extensive simulation results demonstrate that this method maintains high detection and
correction accuracy on wildfire architecture datasets under different pruning ratios. The improved model achieves an
accuracy of 95. 29% at a pruning rate of 80% an increase of 0. 92% effectively addressing the issue of model over-
parameterization. Through various fine-tuning training processes the accuracy of the improved model slightly surpasses
that of the unpruned model while reducing the parameter volume by nearly twenty times. As the pruning rate increases
the model?? s detection performance does not significantly decrease from the original precision level and in some cases it
even slightly exceeds the original model?? s precision. This indicates that there is a significant amount of redundant weights during training. Conclusion This method substantially reduces the model ?? s storage volume while ensuring high
classification accuracy demonstrating practical significance for application in embedded devices with limited neural
network storage and computing capabilities. |
| Key words: deep convolutional network VGG fire detection pruning |
