Research On Shape Analysis Techniques Based On Small-scale Data Learning

Shape analysis including shape description,shape classification and shape retrieval is a hot issue in the field of computer vision.Recently,big data analysis and deep learning have greatly promoted the field of computer vision.However,such positive effects have not extended to the field of shape analysis with the same effect,because the field of shape analysis lacks big labeled datasets.Therefore,in order to further improve the accuracy of the shape analysis algorithms,more and more scholars have begun to pay attention to smallscale data learning.The combination of shape descriptors and semi-supervised learning for shape retrieval and classification is the representative technique.This paper systematically summarizes the key factors affecting the speed and accuracy of shape descriptors,and designs two new shape descriptors based on these factors to enrich the diversity of shape descriptors.The structural principle of the shape retrieval post-processing technique is analyzed in depth,and a novel retrieval post-processing algorithm based on semi-supervised learning is proposed.In addition,this paper tries to use supervised learning based on CNN(Convolutional Neural Networks)in the application of shape classification.The main works and contributions are as follows:1.A Fourier Descriptor based on Multiscale Centroid Contour Distance(FD-MSCCD)is proposed.In a deep analysis on the classical Fourier Descriptor based on Centroid Contour Distance(FD-CCD),this paper finds its description scale is too limited,and then solves this problem.This solution mainly relies on a dynamic centroid point to extract shape feature on different scales.The dynamic centroid point is calculated based on the partial contour.The new descriptor is a simple and fast descriptor,which is suitable for scenarios with speed requirements.2.A Weighted Fourier and Wavelet-like Descriptor based on Inner-Distance Shape Context(IDSC-w FW)is proposed.Inner-Distance shape context(IDSC)is a classical local descriptor with high accuracy and low speed.Global descriptors always have low accuracy and high speed.The proposed descriptor reconstructs IDSC feature,and performs the Fourier transform and wavelet-like transform on it.The results of the spectral transforming are finally combined into a new feature,which has high accuracy and speed.3.A post-processing method based on semi-supervised learning for shape retrieval is proposed,“Online to Offline”(O2O).A fast shape descriptor can create a large speed advantage for a retrieval system.In order to maintain this speed advantage,fast retrieval post-processing algorithms must be used,but there is no fast post-processing algorithms currently.O2 O method can fill this gap.In O2 O method,the post-processing is divided into two stages,offline and online.Its core idea is to transfer the calculation process of high complexity in the online stage to the offline stage.It is suitable for post-processing for fast shape descriptors such as HSC and AP&BAP,thereby ensuring the speed advantage of the entire system.4.A CNN-based model for shape classification is proposed,Shape Net.Shape datasets are usually small(generally less than two thousands images).Small size is an important reason why few researchers use CNN to solve the problem of shape classification.Based on this fact,some data augmentation methods are used to expand the small shape dataset to be suitable for training a small CNN.A small CNN-base network,named Shape Net,is designed specifically for shape classification.The final experimental results show that when the dataset is small(20 images for 1 category),the classification accuracy of Shape Net is slightly lower than the state-of-the-art shape descriptors.When the size of the dataset increases(100 images for 1 category),Shape Net can significantly improve the relative classification accuracy,which can even surpass many state-of-the-art shape descriptors.This paper achieves improvements to previous algorithms in terms of both accuracy and speed.The contents includes two shape descriptors using spectral transform,a shape retrieval post-processing method based on semi-supervised learning and a shape classification network based on CNN.These new algorithms will be useful for improving the real-time application of shape analysis.