| Nowadays,in the high-speed development of intelligent field,object detection algorithm has gradually become an important topic in many computer vision research.The object detection task based on deep learning theory help human so efficiently and quickly find out the object which needs to be located,that we can distinguish the object more accurately in different situations.There are many difficulties in the maintenance of internal boss of modern aeroengine.In terms of the subject of detecting the internal boss of aeroengine,this paper detects the location of target using SSD(Single Shot MultiBox Detector)algorithm model and designs an endoscopic target detection system by combining embedded hardware.this paper firstly uses raspberry pie as the main control unit combining with CSI camera,steering gear and control lever sensor to build a boss data acquisition unit,which can achieve the acquisition process in multi-direction and arbitrary angle.Then it introduces the algorithm theory,network architecture and main outstanding improvements of the algorithm and analyzes the advantages and disadvantages of SSD algorithm model.Considering the special targets to be detected in this paper,The improvements are proposed as follows:1.Through image enhancement strategy and data enhancement strategy to process the acquired boss dataset,so that the various boundary information and details of boss can be well demonstrated and increase the diversity of dataset.It can effectively avoid the overfit and improve the testing accuracy of the model for boss.2.The network architecture of SSD algorithm model is improved using the lower convolution layer to detect the object in the feature map.the size of the feature map is shrinking due to the continuous convolution for the input image,so that the various useful information of object is also reduced.Therefore,the proposed method exploits the lower level feature map to slove the object detection task of the boss.3.The K-means clustering analysis algorithm is proposed to analyze and process the boss training dataset.It modify the size ratio of the generating prediction box in the algorithm model by finding out the internal relationship of the dataset and calculating the appropriate size of the boss marker box. |
