Wavelet Scattering Convolution Neural Network And Its Application In Image Retrieval

In 2012,S.Mallat et al.proposed a novel wavelet transform network structure with translational invariance and deformation stability-wavelet scattering convolutional neural network.This network is supported by strict mathematical theory and has strong robustness.Wiatowski et al.then extended the theory to the wavelet scattering network structure proposed by Mallat et al.and proved that the scale influences the deformation stability,and added the downsampling factor to reduce the scattering feature dimension.Since the invariant scattering network structure has the above excellent properties,it has been widely used in audio and image recognition and other fields.Image retrieval refers to finding a group of pictures that are closest to the target picture from the index database.It has a wide range of commercial application prospects and is a necessary function of the image search engine.The core technology of the image retrieval algorithm is to extract the features of the image.The early classic feature extraction algorithm is relatively poor in accuracy and time-consuming.Although the latest emerging deep neural network works well,there are certain requirements for the scene and training data set.This paper studies the properties and generalization of the wavelet scattering convolutional neural network,and combines it with the similarity measurement method to apply it to the image retrieval field.The specific work is as follows:1.Introduce the relevant background of wavelet scattering convolutional neural network theory;2.The structure and principle of wavelet scattering convolutional neural network are introduced in detail,as well as the related properties of wavelet scattering theory and its theoretical extension,and the related properties of the generalized invariant scattering framework.3.Apply the wavelet-invariant scattering network to the image retrieval field andestablish a reasonable measure of similarity.Experiment with grayscale and color plots of the HSV color space separately.Based on the gabor and molet wavelet framework,a wavelet kernel is designed for each image(this can be input in an arbitrary size),and then the feature coefficients obtained by adding the largest pooled invariant scattering transform are used as the new features.The dimension is greatly reduced,suitable for large-scale image retrieval.