Research On Visual Detection Technology Of Barrels With Energetic Materials Based On Deep Learning

The transfer of barrels with energetic materials is one of the pretreatment processes of energetic materials.There are some problems such as lower efficiency and higher safety risk which current manual transfer method causes.So it is necessary to introduce object detection technology to detect barrels,then guide the industrial robot to grasp and transfer barrels automatically.This paper takes the detection algorithm for barrels with energetic materials as the research object,analyses the characteristics of barrels with energetic materials,and proposes a detection algorithm for barrels based on deep learning after training and evaluating on the self built barrel image data set.The main research contents of this paper are as follows:Firstly,the overall requirements and technical indicators of the detection task for barrels were analyzed,and the preliminary plan of the detection system for barrels was proposed.Combined with the requirements of the detection task,a three-dimensional model of the detection system was established.According to the detection task,the detection accuracy rate reaches 98%,and the detection time is less than 1000 ms,Faster R-CNN was selected as the basic detection algorithm for detection task of barrels after comparing the performance of various detection algorithms.Secondly,a detection algorithm based on improved Faster R-CNN was proposed according to the characteristics of the barrel detection scene.The Res Net-50 convolutional neural network was selected as the backbone network of the improved Faster R-CNN algorithm,the Softer-NMS algorithm was used to screen the redundant generation frame,and the feature pyramid network was introduced to improve the feature expression ability through multi-level feature fusion.In addition,the spatial transformer networks was also introduced to enhance the ability to resist the spatial deformation.Thirdly,the barrel image data set was constructed to train and evaluate the improved Faster R-CNN.The image data set which contains 5850 images was constructed in the order of acquisition,enhancement and annotation.The improved algorithm was trained by transfer learning.The detection performance of the improve algorithm was evaluated by experiments,and the results of experiments shows that the detection accuracy rate reaches99.3% and the detection speed reaches 111 ms,which meets the technical indicators mentioned above.Finally,comparison experiments of detection performance were carried out to verify whether the improved algorithm has a better detection effect in the barrel detection scene.The ablation experiment of four improved measures were carried out,which proved that the detection accuracy of the improved algorithm was 8.6% higher than that of the original Faster R-CNN,through the comparison experiments of different detection algorithms,it was proved that the improved algorithm was better than 94.1% of YOLO v3 and 88.5% of SSD in detection accuracy.In addition,the improved algorithm was applied to the engineering environment.The result shows that improved algorithm in engineering environments can also achieve good detection results.