Key Point Extraction Of Architectural Images Based On Fusion Of Traditional Features And Convolutional Features

In recent years,with the development of neural networks,people have gradually shifted their research focus from text to images.It is precisely because of the emergence of convolutional neural networks that can convert image information into a user-friendly form,making it easier for people to solve problems in the field of images.The foundation of research is to handle the relationship between extracted image features,in order to solve problems in fields such as image recognition and image retrieval.At present,traditional algorithms and convolutional neural networks are used to handle the relationship between images.However,existing convolutional neural network models and traditional algorithms extract image features from different perspectives,and using a single method alone cannot provide a comprehensive feature description of the image.In addition,how to enhance contextual information in the image is also a key issue.In the field of image key point extraction,building images cannot use a single method to improve the accuracy of matching due to the complexity of the information they contain.This article proposes a key point extraction method for building images based on a combination of traditional algorithms and convolutional neural networks,addressing the aforementioned issues between building images.On this basis,attention mechanisms are added to improve accuracy.The research work and main innovative points of this article on key point extraction in architectural images include the following aspects:(1)To increase the number of effective key points in traditional algorithms in the field of building image matching,the SO algorithm is proposed.This algorithm is improved on the basis of the SIFT(Scale Invariant Feature Transform)algorithm and the ORB(Oriented FAST and Rotated BRIEF)algorithm.Combining the SIFT algorithm with the ORB algorithm to extract different key points and improve matching accuracy.Firstly,the process and principles of feature extraction using the SIFT algorithm were studied,as well as the process and principles of feature extraction using the ORB algorithm.The key points and principles of both extraction were analyzed.Secondly,using the same method to describe key points,corresponding feature vectors and rich key point information are obtained,and experimental verification is conducted to compare their accuracy with the original two algorithms.Finally,the ratio testing algorithm is used to effectively eliminate mismatched points,thereby improving the reliability of the matching results.(2)To combine traditional algorithms and convolutional neural networks to increase the number of effective key points in building images,RCS and RCO models were proposed.Firstly,incorporating CBAM(Convolutional Block Attention Module)into the model enhances the model’s ability to extract contextual information,thereby improving the reliability of matching results.Secondly,the key points extracted by the SIFT algorithm and ORB algorithm are used to process the input of the convolutional neural network.The fully connected layer of the convolutional neural network,SIFT algorithm,and ORB algorithm’s feature vectors are normalized and concatenated,enriching the information contained in the key points.Finally,the ratio testing algorithm is used to match key points and eliminate mismatched points,and experimental verification is conducted to compare its accuracy with traditional algorithms and convolutional models.(3)The RCSO model was proposed using the ideas of RCS model and RCO model.Combining the SO algorithm with convolutional neural networks,the input of the model is modified based on the key points extracted by the SO algorithm,and the corresponding feature vectors are used to calculate similarity.At the same time,the RCSO model is designed and conducted detailed experiments on the Hpatches dataset and Brown dataset,ultimately verifying the effectiveness of the RCSO model from multiple perspectives.(4)A prototype system is designed with architectural image de-weighting as the core,and the model is verified to be feasible and effective.