Research On Intelligent Container Code Recognition Algorithm Via Deep Learning

The intelligent identification technology of container code plays an important role in improving the efficiency of railway container transportation and the level of informatization.However,considering the real time requirement under the real scenery and the complex container code characteristics,how to use a lightweight model to achieve the state-of-the-art performance is still a big challenge.In order to solve the above problems,according to the actual situation of the railway container yard,this paper uses deep learning technology to propose an efficient and accurate container number intelligent identification algorithm.The main work is as follows:(1)The article analyze the actual working conditions of container hoisting in railway freight yards so as to design the hardware layout schemes.We also build optical sensors to collect pictures under actual working conditions,and create data sets.(2)The article proposes a pre-process method to cut out the background by counting the frequency of the container code bounding box in the data set,and reduces the picture size when locating code.Using improved YOLOv3 network to design a container code locating network which has fewer parameter and faster speed.Specifically,using Shufflenet V2 as the main body of the network for feature extraction,while reducing the original YOLOv3 network detection branch,reducing model parameters,and designing a network detector based on the K-Means clustering method and analyzing the relationship between theoretical receptive fields and actual receptive fields.Also this article simplifies the loss function of the original network and use local non-maximum suppression algorithm for network post-processing.(3)In container code recognizing part,the algorithm normalize the container code localizing result,extract the code from the high-resolution copy in order to directly recognize the container code sequence.An efficient and accurate image sequence recognition network is proposed,specifically using the residual network.And based on the channel pruning technology,a feature extraction backbone network with strong learning ability and fewer parameters is designed,and a Encoder based on the self-attention mechanism is used to design a sequence recognition with high accuracy and stronger interpretability.The module can well identify the extracted semantic sequence of container box numbers.(4)In training part,this article use warm-up and cosine annealing methods to adjust the network learning rate,and use the generalization ability of the underlying network by pretraining.After experimental verification,the algorithm proposed in this article has a locating accuracy of 99.1% in the container code locating network,which takes 35 ms,and the recognition accuracy of the container code recognition network reaches 98%,which takes 59 ms,and the overall recognition accuracy of the algorithm has reached 97.12% took 94 ms in total,realizing efficient and accurate recognizing of container code.