Research On Industrial Parts Recognition Method Based On Improved DenseNet-BC

Logistics management system plays an important role in enterprises,and warehouse management occupies a core position in logistics management system,so how to build an efficient and intelligent warehouse management system has been the focus of enterprise research.In the current "Internet +" and intelligent manufacturing development background,the concept of intelligent warehousing was put forward.Intelligent warehousing is proposed to effectively solve the rapid classification of industrial parts to achieve storage and acquisition,while the traditional warehouse management system with the increase of the number of parts and parts of the frequent flow will consume a lot of manpower and financial resources.Therefore,how to efficiently manage industrial parts to achieve fast storage and replacement of parts is a key point of intelligent warehousing.In order to solve this problem,this paper combines deep learning to achieve the classification of industrial parts.This paper presents an industrial part identification method based on improved DenseNet-BC.Aiming at the problem of lack of open data set,a complete collection scheme is designed.Since the collected images cannot be directly put into the model for training,this paper proposes a series of pretreatment processes to process the images.Finally,the images are put into the improved model for training,and the performance of the model is proved in the experiment.The main contents of this article are as follows:(1)Due to the low efficiency of collecting a large amount of data manually,a set of automatic acquisition software is developed based on the current acquisition framework,which can automatically collect data,recognize the parts in the camera and train the model in real time.In view of the noise generated in the acquisition process and the differences between individual parts,so in order to extract the key features,this paper optimized the acquisition framework,and finally cooperated with the acquisition software to take the collected images as experimental objects.(2)Aiming at the problem that the resolution of the collected data is too large to be put into the model for training and the quality of the picture,this paper designs a set of process that can automatically cut the picture.First,due to the shooting Angle,parts may not appear in the center of the picture,and secondly,due to the different sizes of parts,they occupy different proportions in the picture.Collected pictures to the cutting may cause only background images,in order to solve the above problems based on the gaussian blur to remove the image noise in order to avoid noise interference localization,then through Canny edge detection algorithm of image to locate the parts,and finally for industrial parts in the picture where the cropping and scaling.Because the noise reduction effect of Gaussian blur depends on the radius of gaussian filter and the selection of radius is affected by the size of parts,a set of process is designed to select the radius.In this paper,the Canny edge detection algorithm is improved,and two direction templates are added to obtain bevel Angle information,so that the edge contour of parts can obtain more information.To solve the problem of small data set,a series of image enhancement operations are carried out,such as random noise increase,image clipping,inversion and so on.(3)An industrial parts classification model based on improved DenseNet was designed and built by PyTorch.Combined with the large number of DenseNet feature channels,the advantage of feature channel redirection in SE Block in SE-ResNet and the special features in parts,the model was optimized and the accuracy was effectively improved.The structure of the model is further optimized by using the focus loss function to solve the problem that the data of some parts are difficult to classify due to the similar shape.Finally,the experimental data show that the improved model has 3.09% improvement in accuracy compared with the original model.