Research On Assembly Line Dynamic Target Detection Algorithm Based On YOLOv3

As a key technology in the field of computer vision,object detection based on deep learning has broad application prospects in the production and processing of industrial assembly lines.As one of the representative deep detection networks,YOLOv3 has achieved excellent detection results on public experimental data sets.However,in the industrial environment,because the targets to be checked are often of various types and sizes,YOLOV3 is prone to miss and mischeck problems of small targets,and it is difficult to deploy and use on industrial equipment with limited resources due to its large number of parameters and calculation amount.In view of the above problems,the thesis improves the YOLOv3 network from the two directions of network optimization and pruning to adapt to the target detection task of Industrial assembly line.The main research contents are as follows:(1)An improved YOLOv3 network based on multi-scale feature fusion is proposed to improve the problem of missed and wrong detection of small targets.Firstly,by deepening the feature extraction network of YOLOv3,4 sets of prediction features of different scales are obtained to form a better multi-scale feature detection network.Secondly,an improved K-means clustering algorithm is used to optimize the dimension of anchor box in the improved network.Then,produced a set of parts data sets containing different sizes and types.Finally,the improved network is trained by using voc2007 and parts data set,and conduct a comparative experiment.The experimental results show that the m AP of the improved network on the VOC2007 data set is 1.3% higher than that of the original network,and the detection speed is 24 fps.The detection accuracy on the part data set is94.2%,and the detection speed is 26 fps,which improves the recognition accuracy while ensuring real-time detection.(2)Propose a channel pruning method for model compression of YOLOv3 network.Methods firstly,regularization is applied to the scaling factor in BN layer to drive the scaling factor to 0 and increase the channel sparsity of YOLOv3 network.Then,the least square method was used to reconstruct the output errors of the convolutional layer before and after pruning.The output errors were further minimized to solve the pruning threshold and complete the pruning.Finally,fine-tune the YOLOv3 pruning network to restore the accuracy of the network temporarily lost due to pruning.The experimental results show that the parameter amount of the YOLOv3 pruning network is reduced by 4.3 times,and the model pruning reaches more than 70%.Significantly reduce the amount of calculation and model space of the YOLOv3 network,and accelerate the speed of network inference.