Adaptive Intelligent Recognition Of"Problematic Map"Based On Multi-Scale Feature Fusion And Convolutional Neural Network

The "problematic map" mainly refers to public maps and map products that endanger national unity,sovereignty,territorial integrity and national security,or other non-compliance with map management regulations.In particular,the "problematic map" referred to in this paper is limited to the current version of the electronic map of China with errors in the following areas: 1)Aksai Chin;2)Southern Tibet;3)South China Sea Islands;4)Diaoyu Dao,Chiwei Yu;5)Taiwan Province.The "problematic map" harms the national interest,seriously violates the seriousness of politics and the rigor of science.It also causes trouble for people to travel and live daily.Therefore,"problematic map" of the investigation and audit is imminent.In this paper,the "problematicatic map" review relies on manual visual discriminant and the popular convolutional neural network requires a large number of training samples.A multiscale feature fusion adaptive "problematic map" detection method based on convolutional neural network in an end-to-end small sample scenario is proposed.For the first time,the algorithm introduces the object detection algorithm based on convolutional neural network into the "problematic map" detection,the multi-scale expression adaptive scale selection mechanism of fusion features is used to realize the intelligent identification and calibration of the five major map errors of "problematicatic map".Further,the method combined with "problematic map" feature to optimize the completion of the "problematic map" smart identification in the case of smaller training samples.Experiments show that the performance is significantly better than the current popular object detection algorithm,which verifies the effectiveness of the proposed method.Specifically,the main research contents and corresponding research results of this paper can be summarized as follows:1)This paper proposes the sample class balance method in the case of unbalanced samples from two aspects of data and algorithm,which solves the situation of class-imbalance well and improves the data quality.At the data level,the "transform oversampling" method is proposed to introduce other information while ensuring the image subject information,which imitates different image representation modes in various scenes of the real world and helps the algorithm to improve the generalization ability.At the algorithm level,the loss function is adjusted based on the ratio of positive and negative samples of the data to achieve class balance.2)This paper proposes a "problematic map" sample mining method that combines the output characteristics of multiple convolutional neural networks.Since there is currently no ready-to-use "problematic map" dataset,this paper research using neural network convolution to mine natural images to obtain a "problematic map" dataset that meets the requirements.For traditional convolutional neural network has a high demand for hardware and a single network can not classify the samples very well.This paper proposes a sample mining method of "problematic map" which fuses the output features of several different convolutional neural networks.This algorithm fuses the features of multiple aspects of the image and can improve the performance of the convolutional neural network while reducing the hardware requirements.3)This paper develops a dataset quality improvement method for the above Chinese map dataset mined by convolutional neural network.The dataset quality improvement method first combines the characteristics that the size of the detection area on different scale maps will change with the scale change,but the aspect ratio of the detection area should not be changed.The label bounding box with different width and height ratios are customized for different detection areas and then the dataset tagging method which is more suitable for the recognition of "problematic map" is obtained.Furthermore,a real-time data augmentation method in small sample scenario is proposed for the dataset of the marked detection area mentioned above.The data augmentation method does not need additional storage space,and not only preserves the original information of the image,but also adds a small amount of additional uncertain information,which greatly enhances the expression ability of the sample and can obviously improve the generalization ability of the model.4)This paper proposes a multi-scale adaptive "problematic map" intelligent recognition method based on convolutional neural network.In this method,the object detection algorithm is introduced for the first time to recognize the "problematic map" and the popular object detection algorithm is improved.The feature pyramid network is integrated into the traditional object detection network and the multi-scale representation mechanism of fusion features is constructed.The fusion features of multiple scales are used to detect the objects in different scales.In addition,the region proposal network was improved to customize Anchor according to the physical properties of the object detection area to obtain region proposal that are more consistent with the size of the detection area.Finally,it is processed after detection to further optimize the detection results.The experimental results show that this method can greatly improve the detection ability of different sizes and sizes of objects in different scales in "problematic map".In particular,this method can significantly improve the detection ability of small size objects.Among them,the average precision used to detect Diaoyu Dao increased from 0.4673 to 0.6710.