改进Inception算法火灾图像识别领域应用

Application of Improved Inception Neural Network in Fire Image Recog

深度学习在智能识别方面具有独特优势,卷积神经网络(CNN)作为其中的典型代表,在火灾图像识别领域的应用得到极大关注。针对经典深度卷积神经网络算法在火灾图像识别应用中存在参数量较大、训练时间长且识别准确率需进一步提高等问题,本文研究给出一种改进Inception算法模型,它以Inception V3为基础,从模型参数、网络结构、卷积通道等3个方面做出改进,增大卷积核以增加感受野,优化网络结构提升模型对火灾图片特征的提取能力。经过在火灾图像数据集的训练及测试,结果表明:改进Inception算法参数量减少到原算法的50%,特征提取效果显著增强;准确率也有所提升,达到97.13%,相比原Inception算法准确率提高了0.90%,相比于Resnet50与Efficientnet算法也有一定优势。

Deep learning has unique advantages in intelligent recognition.As a typical representative of deep learning,convolutional neural network(CNN)has achieved great attention in the field of fire image recognition.The classical depth convolution neural networks have a large number of parameters,long training time and low recognition accuracy.It is in need to be further improved in the specific application of fire image recognition.In this paper,an improved Inception neural network is proposed.Based on the Inception V3 model,the improvements are made from three aspects adjustment:parameters,network structures and convolution channels.The size of the convolution kernels are increased to increase their receptive fields,and the network structure is optimized to improve the extraction ability for fire image features.After training and testing on the flame image data set,the results show that:the number of parameters of the improved Inception neural network is reduced to 50%of the original model.The feature extraction is more significant.And the accuracy is also improved,reaching 97.13%.Compared with the original Inception neural network,the accuracy rate is increased by 0.90%.The improved Inception neural network is also better than the classical models Resnet50 and Efficientnet.