Feature extraction of coal dust particles based on U-Net combined with squeeze and excitation module

To study the inherent mechanism of image characteristics of coal dust particles,a convolutional neural network U-Net combined with squeeze and excitation is proposed for carrying out semantic segmen tation of particle images.The sample data information includes 1 500 coal dust images in six working faces captured from coal preparation plant.Firstly,the image data sets are built with the dust particles by spot sampling and as inputs transmitted to the segmentation network.Secondly,the left-half network is achieved with batch normalization,squeezing and exciting the coal dust features.The obtained features are transferred to the right-half network to up-sample for recovering the image size of particle feature,thus the binary classifi cation of particles is completed.Finally,the binary cross entropy based hinge collaborative loss function is applied to train the network,which can alleviate classification imbalance in segmentation task.The simulation experiments are carried out in three steps as below:① simulating through experiments to evaluate the performance of binary cross entropy based hinge collaborative loss function on network segmentation;② the performance comparison of the proposed network with other conventional image extraction approaches,such as FCN,SegNet,DeepLab,SENet,U-Net network,are evaluated and the feature information extraction ability of Unet SE network is verified;and ③ the eight neighbor algorithm is proposed to locate and label the particles and extract the particle proportion characteristic parameters.The research results show that the binary cross entropy based hinge collaborative loss function can get better segmentation effect,and the improved Unet SE network has a better learning effect on the particle classification with different shapes and sizes.The performance indexes of accuracy,recall rate and F1 scores increase to 0.873 2,0.843 4 and 0.858 0 respectively.It can achieve more accurate extraction,faster training time,and more detailed particles feature acquisition.