单样本学习下时序约束稀疏表示的物体识别方法

Object recognition based on one-shot learning with temporally constrained sparse representation

非接触式传感器相比接触式触觉传感器可以避免与物体直接接触过程中产生的噪声,因而能够获取更有价值的原始数据表征物体内在属性;然而针对非接触式传感器感知的物体属性数据而言,现有算法难以实现单样本学习下的物体准确识别.为解决这一问题,本文提出一种新颖的单样本学习下时序约束稀疏表示方法(one-shot learning with temporally constrained sparse representation,OSLTCSR)用于识别5种不同材料下的50个物体类别.首先将两种原始数据(Lumini光谱和SCiO光谱)并行投影至共享子空间,并且使用聚类典型关联分析法(cluster canonical correlation analysis,C-CCA)计算两种原始数据的聚类相关性特征;其次通过稀疏表示分别计算得到聚类相关性特征数据以及原始数据的编码向量,并利用原始数据的编码向量对相关性特征数据的编码向量进行二次投影映射;然后将两次映射后的原始数据和相关性特征数据进行重构,以充分耦合化两种光谱数据,解决了单样本学习下的数据稀缺问题;进一步地,设计新颖的时序约束稀疏表示方法计算重构后的原始数据和相关性特征数据,以充分考虑每个光谱序列的时序特征;最后与最新的物体识别方法进行实验对比,结果表明提出的OSL-TCSR方法提高了单样本学习情况下的物体识别结果.此外,OSL-TCSR还可灵活迁移至多种应用场景,比如材料识别或纹理识别等.

Compared with the contact haptic sensor,the non-contact sensor can avoid the noise generated during direct contact with the object.Therefore,more valuable haptic data can be obtained.However,it is difficult for existing algorithms to achieve high recognition accuracy under the one-shot learning regime for the data acquired by the non-contact sensor.To address this problem,we propose a novel one-shot learning with a temporally constrained sparse representation(OSL-TCSR)method for recognizing 50 objects from 5 materials.First,two original spectroscopy data items,denoted as Lumini and SCiO,are projected into the shared subspace,and the cluster correlation characteristics of the two spectroscopies are calculated using cluster canonical correlation analysis(C-CCA).Then,we use dictionary learning to calculate the coding vectors of the original data and the correlation feature data after the first projection of two spectroscopies,and the coding vector of the original data is projected into the coding vector of the correlation feature data.Further,we reconstruct the original data and the correlation feature data after two mappings to fully couple the two spectroscopies.Finally,the sparse representation method with temporal constraint is designed to calculate respectively the reconstructed original data and the correlation feature data to consider the temporal characteristics of each spectroscopy sequence.We compare it with the latest object recognition methods to prove that OSL-TCSR improves the accuracy of object recognition in the case of one-shot learning.Moreover,OSL-TCSR can also be applied flexibly to other scenarios,such as material identification or texture recognition.