【作者】 曾磊；

【导师】 董兰芳；

【作者基本信息】 中国科学技术大学 ， 计算机软件与理论， 2020， 硕士

【摘要】 火灾产生的烟雾和火焰对人类的生存环境造成了极大威胁,基于传感器的检测方法存在不经济、延时和误报等问题,而传统的图像处理方法鲁棒性不强,对火灾的预警极为不利。近些年,深度网络模型在视觉领域取得了突出成就,将深度网络模型应用于烟火的识别和检测也成为了研究热点。但是,由于深度网络模型存在高存储和高功耗的弊端,在资源有限环境中对烟火进行识别和检测会受到很大阻碍。为此,本文采用轻量化和通道剪枝方法对模型进行了有效的压缩和加速,主要研究工作可概括如下:(1)本文使用轻量型的网络结构mobilenetv2来精准且快速的识别烟火图像。该网络使用了大量的1*1卷积结构和深度卷积结构来降低模型参数量和计算量,同时又保持了较高的识别准确率。在烟火数据集的实验结果表明,与大型网络vgg16相比,mobilenetv2网络模型的参数量减少了 84.9%,计算量减少了 98.0%,而识别准确率反而高出了 0.76%。(2)本文提出一种改进的refinedet模型来准确和快速的检测烟火图像。该模型将原始refinedet模型中的基础网络vgg16替换成mobilenetv2,额外层缩减为1个卷积层,同时将TCB结构中的卷积模块替换成两个反向残差块模块,从而大幅度地压缩了模型大小并依旧保持了较高的检测精度。在烟火数据集的实验结果表明,相比原始refinedet模型,改进后的模型方法减少了 87.8%的参数量和97.0%的计算量,而mAP值只降低了 0.4%。(3)本文提出一种高效的端到端的通道剪枝方法对烟火识别和检测网络进一步进行压缩和加速。该方法主要是对BN层的缩放因子和平移因子添加L1正则项,并采用FISTA算法来完成对模型的自动剪枝,再一次缩减了模型的大小。在烟火数据集的实验结果表明,对mobilenetv2模型进行剪枝时可以得到40.3%的计算量和36.8%的参数量减少,而分类准确率只损失了 1.72%。对基于改进后的refinedet模型进行剪枝时则获得了43.0%的计算量和47.5%的参数量节省,同时mAP值反而增加了 0.7%。更多还原

【Abstract】 The smoke and flame caused by fire disaster have posed a great threat to the living environment of human beings.Sensor-based detection methods have some problems such as uneconomical,delay and false positives.However,the traditional image processing methods are not robust enough,which are very harmful to the early warning of the fire disaster.In recent years,deep neural network model has made remarkable achievements in the visual domain,and the application of the deep network model in the recognition and detection of smoke and flame has become a research hotspot.However,due to the disadvantages of high storage and high power consumption in the deep network model,the recogonition and detection of smoke and flame in the resource limited environment will be greatly hindered.Therefore,this paper uses lightweight and channel pruning methods to effectively compress and accelerate the model,and the main research work of this paper can be summarized as follows:(1)This paper uses a lightweight network structure mobilenetv2 to accurately and quickly recognize smoke and flame images.The network uses a large number of 1*1 convolution structure and depthwise convolution structure to reduce the amount of model parameters and calculations,while maintaining a high recognition accuracy.The experimental results in the smoke and flame datasets show that compared with the large-scale network VGG16,the mobilenetv2 network model reduces 84.9%parameters and 98.0%FLOPs,while the recognition accuracy is 0.76%higher.(2)This paper proposes an improved refinedet model to accurately and quickly detect smoke and flame images.This model replaces the basic network VGG16 in the original refinedet model with mobilenetv2,and the additional layer is reduced to one convolution layer.At the same time,the convolution module in the TCB structure is replaced by two inverted residual block modules,so as to achieve a large scale of model compression and maintain a high detection accuracy.The experimental results in the smoke and flame datasets show that compared with the original model,the improved model method reduces 87.8%parameters and 97.0%calculations,while the mAP value only decreases 0.4%.(3)This paper proposes an efficient end-to-end channel pruning method to further compress and accelerate the smoke and flame recognition and detection network.This method is mainly to add L1 regular items to the scaling factor and translation factor of the BN layer,and use the FISTA algorithm to complete the automatic pruning of the model,and once again reduce the size of the model.The experimental results in the smoke and flame datasets show that it can prune 40.3%calculations and 36.8%parameters,while the classification accuracy is only lost by 1.72%.When pruning the improved refinedet model,it can save 43.0%calculations and 47.5%parameters,while the mAP value is increased by 0.7%.更多还原