Architectural Style Image Classification Based On Attention Mechanism And Kernel Pseudoinverse Learning

Image classification is an important research direction in the field of image processing.Architectural style image classification,as an application branch of image classification,some research scholars have proposed a lot of architectural style image classification methods.Among them,architectural style image classification based on Convolutional Neural Network(CNN)has received attention from researchers due to its deep network structure,which can extract image features well,but there are still two deficiencies.Firstly,the building is composed of different architectural components,while most methods tend to ignore the weights and spatial relationship between them of different building components when extracting features from building images.Secondly,in the process of using CNN for classifying architectural style images,the training of classification models will cost a lot of computational resources and time with the increase of network depth and width.The attention mechanism can make deep learning more targeted when extracting image features,adaptively learn corresponding weights to different features and extract more discriminative features to make the model’s judgment more accurate.The pseudoinverse learning generally calculates the output weight matrix of the model by matrix inner product and pseudoinverse operation without adopting an iterative optimization process,which is more efficient than the algorithm based on gradient descent.Therefore,to further improve the accuracy and efficiency of architectural style image classification,this paper investigates the architectural style image classification method based on attention mechanism and kernel pseudoinverse learning.The main research work is as follows:(1)An architectural style image classification method based on CNN and channel spatial attention is presented.Firstly,a selection pre-processing operation is added before feature extraction to effectively select the main candidate region of architectural images.Then,the obtained candidate region is used for deep feature extraction by CNN feature extractor.Secondly,a channel spatial attention module is introduced to generate an attention map using channel correlation of different channels of the feature map,which can both assign different weights to different building components and extract the spatial features of different building components.Finally,a Softmax classifier is used to establish the mapping relationship between the attention feature map and the classification labels for predicting the classification labels and their probabilities of architectural style images.Experiments are conducted on Architectural Style Dataset and Architectural Heritage Element Dataset(AHE＿Dataset),and the results show that the method can effectively improve the accuracy of architectural style image classification.(2)To further improve the classification efficiency of architectural style image classification model,a branch neural network(BNN)learning model and classification method based on feature fusion and kernel pseudo-inverse learning is presented.Firstly,to minimize the accuracy loss of the architectural style image classification model,inter-layer feature fusion is performed by using convolutional layers of different sizes to enrich the feature extraction information of architectural style images.Secondly,a BNN trained by Kernel Pseudoinverse Learning(KPIL)is introduced in the network,which allows most of the easily classified architectural images to exit the network early by exiting the threshold classifier with a high confidence level,avoiding learning layer by layer in the baseline network and speeding up the learning efficiency of the network.Finally,coarse to fine image classification is achieved by using shallow-to-deep features of different branch network structures.Experiments are conducted on the Architecture Style Dataset,and the results show that the method can further accelerate the computational efficiency of the network while maintaining a high architectural style image classification.(3)To better apply the proposed methods to practical application,an architectural style image classification system is designed and implemented.