Research Of Shifted Image Recognition Based On Deep Spectrum Histogram Feature

Shifted image recognition is of great importance in the field of image recognition due to its practicability and difficulty.In the real scene,the recognition rate of shifted image is often lower because of the influence of the various factors such as illumination,occlusions,tilts etc.Therefore,shifted image attract more and more attention which results in researching on various kinds of shifted image under the influence of various factors,such as the shifted face recognition,the shifted license plate image recognition,the shifted digital image recognition and so on.In this context,this thesis makes an in-depth study on the shifted image recognition based on the feature of PCANet,which is the representative of the spectrum histogram.The main research contents and contributions of this thesis are summarized as follows:1.Based on the study of the relevant literature on the shifted image recognition,three kinds of image feature extraction methods are analyzed through experiments.The results show that the features of the spectrum histogram are more effective for the recognition of shifted image,which lays a good foundation for the research work of this thesis.2.This thesis explores the influence of the generating modes of network filters and pattern maps on the recognition rate of shifted image recognition based on spectral histogram features represented by PCANet features,and proposes a shifted image recognition network based on Feature Map-Pattern Map-Histogram frame.In the proposed network,the network filters are generated by PCA filter self-convolution and cross-convolution with other filters,which solves the problem that increasing the network filter directly by increasing the convolution layer number of PCANet cannot effectively improve the network's ability to extract the discriminant features.Combining with the dense coding scheme to generate the pattern map features,the network features are affluent and be able to effectively extract the discriminant features of shifted image.Experimental results on AR,LFW,MNIST and other popular benchmark data sets show that the proposed method has good recognition results.3.By analyzing the structure of the shifted image recognition network based on the characteristics of the spectrum histogram,a multi-band feature dimensionality reduction algorithm is proposed from the two aspects of maintaining the network performance and improving the recognition efficiency.Based on the learning of shifted image features and following the basic principle of "sparsity",this algorithm reduces the dimension of multi-band features to ensure the recognition rate and remove the redundant features that affect the network recognition rate,thus the network recognition efficiency is greatly improved.The experimental results on popular benchmark data sets such as AR,LFW,MNIST shows that the proposed dimensionality reduction algorithm not only improves the image recognition efficiency of the network model,but also obtains the recognition rate close to the original network,and even exceeds the original network in the high-dimensional data.