Research On Part Target Detection Method Based On Deep Learning

In the industrial field,with the development of automation and intelligent production,modern chemical plants have higher requirements for intelligent assembly technology.As one of the key technologies of intelligent assembly,part target detection has gradually become a research hotspot.At present,the traditional part target detection algorithm relies on manual feature extraction,which has slow detection speed and low recognition accuracy.In order to solve this problem,this paper takes the target detection of mechanical parts as the research object,and combined with the research and analysis of deep learning technology,respectively selects SSD algorithm and YOLOv3-Tiny algorithm as the basic model to improve,and self-made data sets of mechanical parts are used to verify the improved algorithm.The specific research work is as follows:(1)Aiming at the problem that traditional part target detection algorithms cannot take into account the detection accuracy and detection real-time performance,this paper selects the SSD algorithm that is more balanced in detection speed and accuracy as the basic model,and makes corresponding improvements to it,and proposes a part target detection model MSSD-F.First of all,the traditional SSD algorithm uses VGG as the front-end network.This network has a complex structure and a large amount of parameters,resulting in a large amount of model calculations and slow detection speed.Moreover,the network requires a lot of computing resources and cannot run on embedded devices with limited resources.In response to the above problems,this paper uses the lightweight convolutional neural network Mobile Net V2 instead of VGG as the front-end network to simplify the model structure and reduce the number of parameters,thereby improving the detection speed of the model;secondly,for the negative samples in the training process of the traditional SSD algorithm If the proportion is too large,it will affect the gradient update direction of the loss value to a certain extent,make the model fit prematurely,thereby affecting the problem of model detection effect.This article introduces the Focal loss loss function to replace the original Softmax loss function as the confidence loss function,By adjusting the ratio of positive and negative samples in the training process to optimize the problem of data sample imbalance;finally,a comparative experiment on the self-made mechanical parts data set,the experimental results show that the MSSD-F algorithm compares with the traditional SSD algorithm,the average detection The accuracy is improved by 1.16%,and the detection speed is nearly doubled.From the point of view of the detection effect,the MSSD-F algorithm effectively avoids the missed detection of the traditional SSD algorithm,which verifies the effectiveness of the MSSD-F algorithm.(2)Compared with SSD algorithm,the network structure of YOLOv3-Tiny algorithm is simple and the number of parameters is small.It is a model specially designed for embedded devices with limited resources.Although YOLOv3-Tiny algorithm has better detection speed and can meet the requirements of industrial real-time,the detection accuracy of the model is low because of its simple network structure.Therefore,based on the improvement of YOLOv3-Tiny algorithm,this paper proposes a part target detection algorithm TDS-YOLOv3-Tiny.Firstly,the anchor box parameters are re-clustered through the K-means++ algorithm to obtain the anchor box parameters that are more suitable for the parts of this article;secondly,for the lack of detection accuracy of the YOLOv3-Tiny algorithm,this article improves by increasing the detection scale.The 26×26 size feature map is twice upsampled to 52×52 size,and then combined with the 52×52 shallow feature map output in the backbone network before prediction,which constitutes the third in the model.The experimental results showed that Compared with the YOLOv3-Tiny algorithm,TDS-YOLOv3-Tiny algorithm improves the detection accuracy by 2.79%,and from the detection effect,TDS-YOLOv3-Tiny algorithm improves the detection effect of small target parts.(3)Since there are very few public data sets on mechanical parts at present,a self-made mechanical parts data set is made for this study,and five different types of parts including bolts,nuts,gears,bearings,and washers are selected,in order to improve the data set of parts.Diversity,this article uses a variety of data enhancement methods such as rotation and translation to expand the part data set,and finally the number of samples is increased to 8000.By expanding the data set,not only can the generalization ability of the model be improved,but the possibility of overfitting the model due to insufficient data will also be reduced.