Research And FPGA Implementation Of Pepper Cluster Detection Based On YOLOv5s Model

Pepper is one of the most important cash crops in China,and it ranks first in the world in terms of production and consumption every year.However,due to the pepper trees being covered with spikes and the continuous loss of rural labor,their picking is becoming increasingly difficult.Therefore automatic pepper picking robots have a high prospect of application.To improve the efficiency of automatic picking,fast and accurate detection of pepper clusters is required.FPGA is characterized by high performance,low power consumption,programmability,and parallelized computing methods,so FPGA is chosen as the algorithm inference platform in this thesis.Most of the existing image-based pepper cluster detection methods are traditional methods,and relatively few studies have used deep learning methods to identify pepper clusters.In the process of pepper cluster detection algorithm implementation,there are problems such as lack of data,complex algorithm network structure,poor real-time performance,and low accuracy.In this thesis,to address the above problems,we take mature pepper clusters as the research object and study the improvement of the algorithm model and the deployment of FPGA,the main work is as follows:Ⅰ.Data sets were created by collecting a large number of images of pepper clusters in the fieldThe quality of the dataset is ensured by collecting images at different distances,different light levels and different locations during the ripening time of pepper clusters,expanding the collected images by mirror flipping,rotating,changing brightness,contrast and cropping,and completing the dataset by manual labeling with the LabelImg tool.Ⅱ.A lightweight YOLOv5s pepper cluster detection method based on Shufflenetv2 network is proposed for the pepper cluster detection problemIn order to reduce the number of parameters and computation of YOLOv5s algorithm network model and make it easier to deploy in FPGA,firstly,the algorithm network model structure is significantly reduced by replacing the backbone network of YOLOv5s algorithm with Shufflenetv2 network;then,in order to improve the detection accuracy of pepper clusters,the attention mechanism CBAM module is added at the end of the backbone network and the SPP module is improved to SPPF module;finally,the CIoU_Loss loss function is improved to SIoU_Loss loss function.Ⅲ.Hardware implementation of lightweight YOLOv5s pepper cluster detectionThe FPGA computing platform with Zynq UltraScale+MPSoC architecture is used for parallel accelerated inference of the algorithmic model.The previously improved algorithm is ported to the ZCU104 development board through Vitis AI tool;including the construction of FPGA hardware engineering,quantized.compilation of neural network parameters,and compilation and deployment of Zynq SoC images.The experimental results show that the FPGA platform has superior performance in terms of power consumption,speed,and accuracy in combination compared with other platforms for the inference of algorithmic models.In this thesis,a pepper cluster detection method with high detection accuracy,fast detection speed and low energy consumption is investigated based on the ZCU104 development board of FPGA,using mature pepper clusters in natural environment as the research object,and the YOLOv5s target detection algorithm is lightly improved.