Rust Detection And Application Based On Improved YOLOv3 And Model Lightweighting

With the re-emergence of deep learning,deep neural networks,especially convolutional neural networks,have shown great advantages in computer vision,natural language processing and other fields.They are widely used in image classification,object detection,semantic segmentation,speech signal processing,medical image analysis,remote sensing image analysis and many edge computing scenarios.The main research contents of this paper are as follows:(1)Among the various methods for detecting rust defects on the surface of cranes,manual inspection is inefficient,and the accuracy of some traditional methods with artificial features is poor.YOLOv3 is a fast target detection algorithm that is more suitable for edge computing scenes,but the size of its feature map is too large to detect some medium or large objects.This paper proposes a rust defect detection algorithm based on improved YOLOv3 to detect rust defects on the surface of a crane.A residual block is added to the YOLOv3 network to extract smaller-scale features.Then perform a down-sampling operation on the features extracted by the 4th,5th,and 6th residual modules in the feature extraction network,and fuse them with the features extracted by the 3rd,4th,and 5th residual networks respectively to obtain the features with stronger characterization capabilities.This method achieves 92% detection accuracy on the crane rust dataset,which is better than the initial YOLOv3 network.(2)The deep learning model cannot be run in real time on the edge computing platform with limited computing resources due to its large amount of calculation.To solve this problem,this paper proposes an indirect model lightweighting method based on the importance of the convolutional layer channels.A scaling factor is attached to each channel as a parameter basis for measuring the importance of them.Through sparse training,the value of the scaling factor corresponding to the channel with lower importance is approached to 0,and then these channels and their corresponding input and output connections will be pruned from the network.The direct effect of this method is to reduce calculations of one model and the storage space occupied by it,and the speed of model inference will be indirectly increased meantime.When the pruning ratio of the parameters reaches0.85,the model inference speed is doubled,while the detection accuracy is only reduced by 0.15%.(3)The indirect model lightweighting method reduces the complexity of the model,but it cannot directly and quantitatively limit the model's energy consumption,inference delay and other computing resources.The resulting lightweighting model of the indirect method still cannot run in real time on the edge computing platform(such as Jetson TX2).In order to further improve the inference speed,this paper proposes a direct model lightweighting method based on delay constraint.This method takes the inference delay as the constraint during model training,and transforms the model training process into the process of solving optimization problem.The optimal solution is obtained through the ADMM optimization algorithm,and a model with a shorter reasoning delay can be directly obtained in the meantime.To ensure the model accuracy,the lightweight model obtained by the direct model lightweighting method can reach up to 4 times the inference speed of the original model;when the accuracy drops by 10%,the inference speed can reach 5 times the original model.