Building Identification Based On Deep Learning

With the rapid development of urbanization,the identification between cities is becoming weaker and weaker,and people have already felt visual fatigue to the cookiecutter urban buildings.Those buildings rich in historical and cultural deposits have become the target for all cities to imitate,and all cities begin to build some buildings with different characteristics.These buildings are not only the landmarks of the city,but also the inheritance of the city's history and culture,as well as the coordinates of urban orientation,and also an effective means to attract tourists.And multifarious characteristics of building the recognition of people has brought great influence,deep learning approach in the field of image recognition in recent years has been a huge success,it has a strong ability to learn and efficient characteristics,the advantages of more important from the raw data to the abstract semantic concept of pixel level can extract information step by step,which makes it the global feature extraction has many advantages,compared with the traditional pattern recognition has a higher recognition efficiency.In this thesis,based on the deep learning method,research,design and implement a set of identification method of the building,Let people find the location of these interesting "popular buildings" and bring potential tourism value to these cities.This thesis makes a special data set of landmark buildings according to various shooting environments of buildings.The data set contains samples from different weather,different light and different angles,effectively improving the model's fit resistance.In this thesis,the Faster R-CNN algorithm is used as the basic training model to improve the original basic network and adopt a dense-connected residual block network DRNet for special scenes of building identification(such as buildings being blocked,situation at night,etc.).This kind of network can use the previous feature block diagram and the output result of this layer to superimpose on the channel dimension to achieve the effect of feature reuse.The resulting feature block diagram not only does not lose the low-level edge texture information,but also multiplexes the low-level feature block diagram in the deep convolutional network,which makes the fused feature block diagram have more abundant feature information and effectively improves the recognition rate of the model for photos taken in complex environment.Due to the different image sizes of the collected data set,and the use of two integer quantization in the RoI Pooling layer to extract the feature block diagram of the original model,there is a certain difference between the actual candidate box and the obtained candidate box,and the feature block diagram has a certain degree of deformation,resulting in the loss of spatial information of the image and the reduction of the accuracy of the feature block diagram.Therefore,this thesis adopts the new RoI Align layer extraction feature block diagram,and solves the problem of regional mismatch in the original algorithm by bilinear interpolation.The experimental results show that the proposed method can achieve 82.1% mAP for landmark buildings and the prediction of building coordinates is more accurate,under the condition that the training data set is sufficient.Compared with other models,the model not only has a good recognition effect on normal images,but also has excellent performance in the recognition of images captured in complex environments.