Research On Detection Algorithm Of Electrical Components For Educational Experiment Platform

Object detection is one of the hottest research topics in the field of artificial intelligence,and it has a wide range of applications.In the physics teaching circuit system experiments of traditional middle school students,it is necessary to evaluate and score the circuit system connected by the students.In the teaching process,the numbers of teachers is limited.Due to the large number of students,it is impossible to score each student's experimental results in real time.If the students are scored sequentially,it will cause a great waste of time.Therefore,in order to make real-time scoring of student experiment results,reduce the teaching time and improve scoring efficiency,it is very necessary to develop a real-time online student physics experiment scoring platform.In the embedded platform,the most important thing for real-time scoring of physics experiments is the real-time detection of electrical components.If the detection speed is slow,it will still slow down the teaching speed and affect the students' learning.In traditional detection algorithms,Using traditional detection algorithms to detect electrical components,for example,use the HOG+SVM algorithm for detection,Although it has the characteristics of high detection rate and low false detection rate,the detection time will be very long and can not meet the real-time requirements,In order to solve the problem.Think from the actually,a system for rapid detection of electrical components is proposed.The SSD(Single Shot Multi Box Detector)object detection algorithm is used as the main object detection algorithm of the electrical component detection system in this article.Because this algorithm is not friendly to small target detection,this topic propose one block which use the different net feature to constitute,this block can improve the accuracy about the small thing detection.In the traditional NMS target detection algorithm,due to the high overlap of similar targets,the pre-selection box is prone to misdetection or missed detection of the target object during the elimination process.A new Re Distance Io U-NMS algorithm is proposed in this paper to replace the NMS algorithm to improve the detection accuracy of similar targets with excessive overlap.In the embedded development process,in order to further improve the model performance,the embedded development board used in this article is the RK3399 Pro development board based on the NPU processor.To further improve the model in the embedded device running speed.This article draws on the idea of Google's Mobile Net network to replace the traditional convolution part of the SSD object detection algorithm with a deep separable convolution.In the process of model transplantation,the int8 offline hybrid quantified method is used to further compress the model trained on the PC side in this article to obtain a high-precision,low-parameter network model.The experimental results show that the SSD algorithm with hierarchical feature fusion module proposed in this paper has higher detection accuracy than YOLO,Faster R-CNN and other series of detection algorithms in small target detection results.Compared with the original SSD target detection algorithm,replacing the traditional NMS algorithm with Re Distance-Io U NMS has higher detection accuracy in the detection of similar objects.At the same time,the algorithm proposed in this paper can realize fast target detection in the development version of RK3399 Pro.Using the int8 type offline hybrid quantized network model,the model compression rate reaches more than 70%,while the detection accuracy in the embedded device reaches 77.4%,and the detection speed reaches 25.6FPS,which can be more efficient and accurate in practical applications.