基于多尺度分组卷积ResNet34的岩石识别模型

Rock recognition based on multi-scale grouped convolution ResNet34

岩石分类有效提升地质风险评估的效率,为提高岩石粒度特征的识别精度,提出基于多尺度分组卷积ResNet34网络的岩石识别方法,在残差模块中添加多尺度分组卷积和空洞卷积,提升网络的特征表达能力.采用多尺度分组卷积特征提取的方式,将特征图按通道方向分为四部分,分别采用不同大小的卷积核进行并行运算和拼接,在更细粒度上提取岩石特征,同时减少了模型训练参数,进而采取空洞卷积增加感受野大小,提升岩石粒度识别精度.实验表明,该方法不仅有效提升了网络训练的收敛速度,而且在2021年第九届“泰迪杯”数据挖掘挑战赛七类岩石数据集上识别准确率达到97.6%.

Rock classification effectively improves the efficiency of geological risk assessment.In order to improve the recognition accuracy of rock particle size characteristics,a rock identification method based on the multi-scale grouped convolution ResNet34 network is proposed,and multi-scale grouped convolution and cavity convolution are added to the residual module to improve the feature expression ability of the network.Using multi-scale grouped convolution feature extraction,the feature map is divided into four parts according to the channel direction,and convolution kernels of different sizes are used for parallel operation and splicing,and rock features are extracted at a finer granularity.At the same time,the model training parameters are reduced;then the cavity convolution is used to increase the size of the receptive field and improve the accuracy of rock particle size recognition.Experiments show that this method not only effectively improves the convergence speed of network training,but also achieves a recognition accuracy of 97.6%on the seven types of rocks for the 9th"Teddy Cup"Data Mining Challenge in 2021.