Research On Building Recognition Technology Based On Transfer Learning

Today’s buildings are evolving toward diversification of styles and functions.As an important branch under scene image recognition building recognition is also a very important and challenging research content in the field of computer vision and robotics.In this way how to use computers Thinking makes it possible to automatically understand building images and effectively identify them,and take the opportunity to serve image retrieval in a big data environment,which has become a problem that needs to be solved urgently.A key technology of building recognition is feature extraction,but traditional features have very large limitations.At the same time artificial feature engineering is time-consuming and labor-intensive and requires high professional knowledge.The extracted features are relatively single,and with the complexity of the data set with the increase,the redundant information in the image will become more and more numerous,and the use of traditional machine learning methods cannot meet the needs of image feature extraction.In recent years,with the rapid development of deep learning technology.Based on the advantages of big data and high-dimensional parameter space,the end-to-end neural network structure gradually abstracts and synthesizes high-level features from the bottom to the top.The data-driven self-learning method ensures that the convolutional neural network has excellent feature extraction capabilities.Therefore,deep learning and the combination of building identification has practical application value.To this end,this paper mainly completed the following aspects:1.At first the paper introduces the relevant work of building recognition,Including explaining some commonly used methods and technologies at home and abroad,and analyze the difficulties existing in the current building image category.The common methods used in the key steps are analyzed and summarized.2.Aiming at the problem of traditional feature extraction with low performance and poor versatility,this paper proposes a Recog-Net image authentication model based on transfer learning convolutional neural network.The features extracted by the network can contain rich scene semantic information with Inception-Net as the feature extractor.Meanwhile the model training uses the "pre-training-fine-tuning mode" based on transfer learning.Through Image Net the feature extractor is pre-trained and the part from the convolutional network to the bottleneck layer is used as the feature extraction process.In order to allow the model to better capture the potential correlation information between features to improve the prediction accuracy of the model,a multi-feature calibration technique is proposed.By artificially defining the multi-functional buildings,the bottleneck is passed the feature vectors after the layer are more representative and can be applied to small sample data sets.In the end,experiments show that the features obtained by multi-feature calibration have better performance than traditional features and other convolutional neural network structure learned features,with high versatility,recognition rate,and robustness.The recognition effect of the image has a very good improvement effect.3.Considering the distribution of feature information in the space of different building pictures is not the same,a building image often contains a wealth of building image feature information,and at the same time is mixed with irrelevant background information,so how to choose a building image Extracting regions with rich key information and ignoring irrelevant secondary regions are very important for improving the accuracy of building image recognition.Therefore,this paper proposes a multi-angle salient area building image recognition scheme.Transform the size and scale of the building picture,and multi-scale crop the building image to obtain the location of the region with rich semantic information(for example,the red cross feature obtained by cropping the hospital picture can greatly improve the hospital recognition effect)The image information(such as flowers,trees,etc.in the picture)is discarded,and then the multi-scale picture is sent to Recog-Net.To form multi-angle features,and the classifier is replaced by SVM for classification.The experiment shows that the image features of the multi-angle salient area Extraction has a great influence on the improvement of image recognition rate.