Small Object Detection And Character Recognition Based On Deep Neural Network Attention Model

Object detection is to locate and classify object from image or video.The task of detecting small object with a little of pixels is called small target detection.Small targets have fewer pixels and less information so that these features are easy to be drown in the background.Thus,small target detection is always the difficulty of target detection.The task of character recognition is to extract characters that exist in the image and combine them into string output.The performance of character recognition in natural scene is poor because of the complex background,the variety of fonts and the random distortion of characters.In overhead catenary system(OCS),pillar number plate uniquely identifies the actual location of image acquisition.OCS pillar number plate detection and recognition are usually divided into two sub-tasks,forming a sequential pipeline: number plate detection and number recognition.In pillar number plate detection task,OCS images might be distorted,blurry and small due to the relative movement and camera focal.If the shooting distance is far from target,pillar number plates could occupy a little area which makes the pillar number plate detection a hard work.In pillar number recognition task,images have a low resolution and the image qualities are always poor.In order to solve these limitations above,the main works of this article are as follows:This paper put forward the Attention-guided Multi-Scale CNN for pillar number plate detection.Based on YOLOv2,our method designs an attention module that captures important features in a complex scene.Additionally,multi-scale feature fusion strategy is designed to improve the performance of small target detection.In the attention model,a pair of up-down sampling is used to enlarge the receptive field.The attention module uses skip connection strategy to connects the attention mask with a deeper feature map,aiming to improve the feature extraction ability of our method.In multi-scale feature fusion stage,our method integrates the deep feature map with the shallow feature map.The results will be gained by detecting the fusion feature map.This strategy could improve the detection accuracy of the small target.The experimental results show that our method could keep a good detection accuracy in all kinds of complex OCS images.This paper designs the pillar number plate recognition method which is based on the feature reorganization attention module.Firstly,the convolutional neural network is used to extract the feature of the license plate images.The features are entered into the encoder-decoder part to obtain the possibility of each character in each sequence.Finally,the decoding results are put into the transcription layer to get the final outputs.We put the attention module into the encoder-decoder part,which is responsible for feature reorganization.We try different feature scale adjustment approaches and different calculation formulas of attention model based on the analysis of the different connection ways.Our method obtains 97.1078% recognition accuracy on the OCS pillar number plate recognition dataset.Experiments prove that our method has high recognition accuracy rate and robustness.3.This paper designs and realizes the automatic pillar number plate detection and recognition in OCS.According to the actual needs of users,our system designs the user basic information table and image basic information table for managing.We also propose three modules: user management module,data management module and algorithm execution module.The system realized in this paper is easy to use and has good human-computer interaction.