The Container Character Recognition Based On Deep Learning

The container is an important bulk logistics transportation method.Automatic recognition of characters can accelerate the realization of information and automation in the port.This paper proposes a set of localization and segmentation methods and a deep learning-based recognition method for container character segmentation and recognition problems.The automatic recognition of character 1s achieved and a high recognition rate is achieved.The specific works are as follows:1)Based on the analysis of the characteristics of actual container images,the thesis proposes a preprocessing method based on color features.This method selects the appropriate threshold values for the difference between the background color and a non-background color and two non-background component color maps respectively binarize and then perform an AND operation to obtain a high-clarity binarization map containing the target region.Compared with conventional methods,this method overcomes the effects of external lighting and bumps,eliminates most of the interferences such as advertisements and smudges,and lays a good foundation for subsequent positioning and segmentation.2)The thesis designs a method for locating and segmenting characters based on two projections,and uses Radon transform to correct the characters region.The method firstly uses horizontal projection to coarsely locate the container characters area,and excludes other areas in combination with the area of the characters area,the number of characters,etc.to achieve accurate positioning of the characters area.Then Radon transform is used to make an inclination correction to the characters area.Finally,the horizontal and vertical projections are combined to determine the position of each character and segment it out.This method is simple and has good real-time performance.3)Research deep learning and LeNet-5 network,and improve it.The improved LeNet-5 model consists of six layers.The network uses the ReLU activation layer and joins dropout layer and changes the pool2 layer to the spp layer.This reduces the effects of model parameters and gradient disappearance.4)Take advantage of the actual container images and data extensions to form larger scale data sets to train and test the improved LeNet-5 network.Experimental results show that the recognition rate of this method reaches 96.79%,the recognition time of single character is close,and the recognition speed on ordinary CPU is about 1s.Compared with BP neural network,template matching and feature matching methods,it is found that the improved LeNet-5 network has higher superiority in recognition rate.