An Automatic Container-code Recogntion System Based On Neural Networks

With the development of global trade,the number of containers to be shipped is increasing year by year.Because of the huge amount of containers,management of containers is essentially required by the shipping industry.Most of the management is done by manually recording the container-code.While this way of management requires a huge cost.Therefore an aumatic container-code recognition system is required immediately.There are many container-code recognition system already.Some of them are called radio-frequency technique,which requires additional equipments for container-code recognition,which will increase cost in the container-code recognition system.Some of the techniques are based on vision,and this method requires no additional equipment,therefore it is much more cheap than the radio-frequency technique.But the technique of conatiner-code recognition system based on vision are more difficult.An(ACCRS)automatic container-code recognition system based on vision is proposed.This system not only save the cost spent on ACCRS,but also provide a efficient and accurate container-code recognition.This system consists of two main modules,one is the container-code detection module,and the other is the container-code recognition module.The detection module utilizes both detection techniques based on traditional computer vision and techniques based on neural network.Techniques based on traditional computer vision output detection results with high precision,and container-code's position in image can be determined by the detection results from the techniques based on traditional computer vision.Techniques based on neural network are robust under disturbance,such as invariance of illumination,weak illumination,incomplete container-code,and image blue.A module used for synthesis is proposed in this system,which is responsible for combining detection results from computer vision and neural network.Through combination,the system is able to generate detection result with both hight robustness and high precision.The recognition module in the system utilize two recognition techniques.One is called segmentation-based recognition,which will recognize each character in container-code one by one.And the other technique called end-to-end recognition,which takes the whole container-code as its input,and output a sequence of characters,which is the predicted container-code by the end-to-end recognition module.A module used for combining both recognition results also exists in recognition module.The combination is based on the verification of conatiner-code,and based on a reasonable combination scheme,the system is able to generate an accurate container-code recognition result throught combination.Through experiment,the automatic container-code recognition system can achieve92.2% of recogntion accuracy throught combination of the two recognition technique,which is 6% higher than the accuracy of the segmentation-based method,and 13% higher than the accuracy of the end-to-end based method.