Design Of Software System For Detection Of Circuit Experiment Equipment Based On OPYOLOv3-Tiny Algorithm

In recent years,the technology of object detection algorithm based on Deep Convolutional Neural Networks and embedded platform has been widely used in many fields.However,in domestic education field,the related applications are almost only used in the face detection system of examinations in the domestic education field,which actually didn't apply to the primary teaching work deeply.Based on the research of physical circuit experiment in the junior high school,YOLOv3-Tiny object detection algorithm and the embedded platform RK3399 Pro,this thesis desings a software system that can replace teachers and students to complete detection before the beginning of the circuit experiment,which greatly improve the experimental efficiency.The main content of this thesis includes:(1)Compare with the current mainstream object detection algorithm,this thesis explains these detection principles,the pros and cons of each algorithm in the real detection,and introduces several common model compression methods.(2)Analyze the detection process and result of YOLOv3-Tiny algorithm in detail,and apply the method of overlapping Max Pooling to the algorithm's Max Pooling layers to effectively reduce the loss of characteristics.At the same time,this thesis focuses on multi-feature fusion method to make full use of the output parameters of the shallow Conv layer to add the third feature map to the algorithm's prediction forecast network,which can improve the detection effect of small circuit experimental equipments.Finally,the K-means clustering algorithm with improved distance formula is used to perform the best Anchor clustering on the dataset of homemade circuit experimental equipment and optimize the algorithm parameters.In the GPU(Graphic Processing Unit)environment connected with RK3399 Pro,the improved OPYOLOv3-Tiny algorithm has achieved great results in terms of detection accuracy and speed.(3)Analyze the necessity of the conversion of the framework on algorithm model and detailed using methods of conversion tool — RKNN-Toolkit.Due to RK3399Pro's Random Access Memory is too small and computing resource is so limitled,the actual detection frame rate on the embedded platform has a sharp decrease.To solve this problem,this thesis has put forward hybrid and fixed-pointed quantization methods to compress the OPYOLOv3-Tiny algorithm's model files so as to reduce the model size by 68.8% while ensuring the accuracy of the model.(4)Based on the hybrid and fixed-pointed quanfication model,a real-time detection system is designed for circuit experiment equipments under Android 8.1 of embedded platform RK3399 Pro,which can run automatically.The model reasoning module is also used to call the underlying NPU(Network Processing Unit)through RKNN-API function in the software system so as to realize the acceleration of NPU and real-time detection of circuit experiments.A detection system of circuit experiment equipment based on the embedded platform is designed in this thesis to keep a balance between the speed and accuracy,and finally realizes the low-cost,and high-real-time object detection system,which lays a solid foundation for the teaching of physical circuit experiment in the junior high school.