Application And Realization Of Few Shot Learning In Robot Vision System

With the unremitting efforts of researchers,deep learning-based machine vision systems have reached benchmarks close to or even beyond humans for many classification tasks.However,when the robot vision system solves the classification problem in the case where the labeled data is scarce,the performance often deteriorates because the model is overfitted.Therefore,Few-shot learning emerges a new way to solve the problem in deep learning.Few-shot learning is able to quickly learn key information from only a small number of reference samples,so as to be qualified for a new classification task.In addition to solving the problem of deep learning under conditions of data scarcity,few-shot learning is also instructive for exploring general artificial intelligence.By introducing the representative differences of reference samples into the distance measurement method,the model obtains a better class representation.This paper proposes a simple and efficient few-shot learning network,namely Critical Drection Projection network.Basically,two crucial steps are involved in CDP: The first step is to find the critical directions for each category in embedding space,and the second step is to measure the similarity between samples and critical directions according to projection length.The main contributions of this paper include:1)In order to improve the accuracy of few-shot learning in classification tasks,this paper proposes an efficient and simple few-shot learning network,Critical Direction Projection(CDP)network.CDP improves the network from multiple aspects such as feature extraction,distance measurement,training strategy,etc.Compared with other networks,the CDP network reaches the state of art in multiple benchmark dataset experiments for few-shot learning.2)In order to solve the problem of difficulty and high cost of 3D data acquisition,this paper extends critical direction projection network to few-shot learning of 3D point cloud data and proposes a new few-shot learning dataset Few-shot Model Net40 based on Model Net40,which provides a good reference for future research.3)In order to solve the problem of indoor scene recognition when the task target scene category is unknown,and improve the adaptability of robot to complex environment,this paper applies critical direction projection network on indoor scene recognition problem.The benchmark datasets and artificially collected datasets for indoor scene recognition tests are performed.The test results show the potential of few-shot learning for scene recognition problems.The critical direction projection network solves the problem of scene recognition when the target scene is unknown,and improves the adaptability of robot to complex environment.