Research On Recognition Algorithm Of Coal Gangue Image Based On Deep Learning And FPGA Implementation

With the continuous promotion of the national dual carbon industrial tasks of "carbon peak","carbon neutralization" and the wave of Industry 4.0,coal preparation,as an important link in the clean utilization of coal resources,has gradually moved from automation and information to intelligence.Gangue is the main solid waste produced in the process of coal mining,so how to identify and sort the gangue cleanly and efficiently is particularly important.At present,compared with traditional image recognition methods,coal gangue recognition method based on depth learning can automatically extract target features,and the recognition accuracy is higher.therefore,more and more scholars study this method.However,most deep learning algorithms have a large number of parameters and are complex to calculate.They need to run on high-performance CPU or GPU hardware devices,and the actual application scenarios are severely limited.To solve this problem,this paper improves the algorithm and hardware.At the algorithm level,this paper selects the lightweight version of the current target detection algorithm YOLOV4,YOLOV4,and improves its model on the basis of the original algorithm.This paper uses the idea of packet convolution to design two lightweight feature extraction modules S-Block1 and S-Block2,replacing the first two Resblocks in the original backbone network with S-Block1＿Body module,replace the last Darknet Conv2 D with S-Block2＿BN＿In the Leaky module,the parameter amount and calculation amount of YOLOv4 tiny algorithm have been reduced.On the hardware level,this paper uses FPGA as the hardware implementation platform and Xilinx’s Zynq-7020 as the FPGA chip.The algorithm hardware deployment process is relatively complex.First,BN layer fusion is carried out for the trained algorithm model weights to reduce the model weight parameters and computation;Because the algorithm includes a large number of convolution and pooling operations,this paper designs a convolution and pooling IP core to speed up the operation.In the IP core design,a variety of optimization methods are used,including 16 bit fixed-point quantization,pipeline operation optimization,table tennis operation and parallel operation of input and output channels;Finally,the whole hardware system and software system are designed.The experimental test was carried out in the self-made coal gangue data set.The experimental results show that in the computer side experiment,the improved YOLOv4-tiny algorithm is used to recognize coal gangue,with a value of 97.12%.The detection time of each image is 53 milliseconds,which greatly improves the recognition speed while ensuring high detection accuracy.In the FPGA platform experiment,the detection accuracy is slightly lower than that of GPU and CPU,the value of m AP is 96.56%,the recognition speed of each image is 346 milliseconds,the hardware power consumption is only 2.855 W,and the energy consumption ratio is 3.1times and 2.9 times of that of CPU and GPU respectively.Figure 54 Table 6 Reference 83...