基于视觉选择性的离变焦图像序列慢变特征提取算法研究

SLOW VARIATION FEATURE EXTRACTION ALGORITHM FOR DEFOCUSED IMAGE SEQUENCES BASED ON VISUAL SELECTIVITY

传统慢特征分析(SFA)方法提取的慢变特征不能揭示自然图像的视觉空间拓扑结构。基于此,提出基于视觉选择性的离变焦图像序列慢变特征提取算法。以myTICA方法替代源SFA算法的PCA方法,提取能够反映自然图像离变焦图像序列的视觉空间基的种类、每类元素数量、类内与类间拓扑结构不变性的Gabor特征,并建立与该序列对应的不变性特征森林;利用蒙特卡洛马尔可夫(MCMC)算法替代源SFA算法的多项式扩张方法,实现Gabor类内的元素扩展;利用自定义的近似正交剪枝算法实现不变性特征森林的优化,顺次解决元素法的采样丢失与森林优化问题;利用自定义响应度计算规则实现特征匹配。实验结果表明:该算法正确可行,具有较好的抗噪能力;在实验阈值为0.4时,算法获得识别率为99.96%,说明该算法具有较强的分类能力。

The slow variation features extracted by the traditional slow feature analysis(SFA)cannot reveal the visual spatial topology of natural images.Aiming at this problem,this paper proposes the slow variation feature extraction algorithm for defocused image sequences based on visual selectivity.We replaced PCA method of the original SFA algorithm with myTICA method,and extracted Gabor characteristics of natural image zoom image sequence that could reflect the class of visual space base,the number of elements in each category,the invariance of intra-class and inter-class topological structure.The invariant characteristic forest corresponding to the sequence was established.Using the Monte Carlo Markov(MCMC)algorithm instead of the polynomial extension method of the original SFA algorithm,the element expansion in Gabor class was realized.The self-defined approximate orthogonal pruning algorithm was used to optimize the invariant characteristic forest,and the problems of sampling loss and forest optimization of element method were solved.We used custom responsiveness calculation rules to achieve feature matching.The experimental results show that the algorithm is correct and feasible,and has better anti-noise ability.When the experimental threshold is 0.4,the recognition rate of the algorithm is 99.96%,which shows that the algorithm has strong classification ability.