Research And Implementation Of Image Recognition And Localization Based On Deep Learning

With the deepening of computer vision research,researchers began to expand their application in the industrial field,and object recognition and localization is a crucial link in computer vision.As one of the application fields of computer vision,the smelting magnesium reduction process has higher requirements on the accuracy and real-time performance of the object detection.However,traditional methods are difficult to meet the requirements of computer vision in this field of accuracy and detection speed.The Deep Learning convolutional neural network has the characteristics of autonomous learning image features,so it has a greater advantage in computer vision than other traditional methods.Therefore,based on Deep Learning,this paper describes and analyzes the object recognition and localization and its application in the process of magnesium reduction..In this paper,the representative algorithms of target detection,Faster RCNN and YOLO,are the research focus.The feasibility of the two algorithms is verified in the smelting magnesium industrial environment,and the advantages and disadvantages of the two algorithms are compared.At the same time,the algorithm is optimized based on the problems generated by the experiment.For the problem of imperfect industrial data collection,this paper uses transfer learning and data expansion,solved the problem that the detection accuracy caused by insufficient data set is not up to standard.For the multi-scale information problem existing in the data set,according to the anchors generated by the prior algorithm and the optimized K-means clustering algorithm,compare and analyze the AP value and IOU relationship of the network structure,the experiment verifies that the anchors generated by K-means clustering isbetter than the priori anchors.At the same time,for the problem that the detection accuracy is insufficient when there are multi-scale information in the picture,the optimized anchors can effectively improve the detection accuracy.Secondly,the region proposal network and the optimized network are compared to the quality of candidate region extraction,it is verified that the quality of the improved network extraction candidate area is better than the original region proposal network.Finally,the non-maximum suppression algorithm is optimized,and the optimized non-maximum suppression algorithm can effectively reduce the false detection.The experimental results show that the optimization operation on the Faster RCNN can accurately and efficiently detect the tank lid and the workers,and at the same time perform robustness tests on the public data set to verify the universality of the network.This paper realizes the recognition and localization of images.On the one hand,the object detection algorithm of deep learning is realized under the Caffe,and the construction of the detection system is completed.In addition,for the acquisition of positioning information,the Kinect is calibrated in the ROS environment to achieve position information acquisition within a small error range,and the positioning system is completed.