Research On Target Detection And Application Based On Improved YOLOv3

With the development and application of machine vision,the application of visual inspection technology in industrial scenarios has become a boom in intelligent manufacturing.In the production process of molten steel,there are many uncontrollable quality problems,and some defects such as scratches and deformation appear during the manufacturing process.The operators need to judge the quality of the pouring steel claws closely.For this reason,this article analyzes and applies deep learning-based target detection framework to train and detect,for achieving different shapes and multi-angle steel claw recognition and detection.The thesis mainly completes the following work:(1)Analyzes the actual environment and requirements of detection targets in this article,and introduces the significance of the introduction of machine vision in the industrial intelligent manufacturing industry and the current situation of mainstream target detection algorithms at home and abroad.(2)According to the development of target detection algorithms,from the two directions of traditional detection algorithms and detection methods based on deep learning,the classic target detection algorithms in the field and their advantages and disadvantages are briefly introduced.(3)According to the environment and requirements of the detection object in this article,if using the sliding window algorithm to find the corresponding position of the steel claw for extracting feature information,sending them to a classifier such as SVM to achieve feature recognition and classification of the target.Under the environment,the requirements of detection accuracy and real-time detection cannot be guaranteed.At the same time,the algorithm has the disadvantages of redundancy and high complexity.Select SSD,Faster-RCNN,and YOLO based on deep learning to train the collected data set,and select a better algorithm to further explore and try.(4)For the detection accuracy,especially for small targets,the detection effect is poor,the dual-channel attention model CBAM is introduced to further optimize the incoming feature information of the prediction layer and improve the detection accuracy;adding a prediction layer from the perspective of network structure can improve the detection efficiency of detecting small target objects better,update the information of the detection object dataset in this article;introduce the modification of perfect pairs in the direction of the loss function to achieve the function of improving the detection effect.(5)Aiming at the improved target detection algorithm in this paper,a graphic user interface application based on implementation is designed to complete the function of one-click target detection,and related basic image preprocessing operation function keys are added.