Biological Motivated Image Features Classification Based On SMF Model

Humans and primates outperform the best machine vision systems withrespect to almost any measure. Thus, people have paid attention to biologicalinspired model since1950s. In1999, Riesenhuber and Poggio proposedSMFs (Standard Model Features) based on the visual perception mechanism.However, there’re few of large-scale evaluation of object recognition andtexture classification based on SMFs from the view of computer vision,especially comparing with SIFT, LBP features etc. popular computer visionmodels, which is of important theoretical and application value for theresearch of biological motivated feature extraction. We carry out quantitativeanalysis and evaluation of selectivity based on selectivity index methodproposed by Leibo.We carry out quantitative analysis and evaluation of rotation, scale andaffine invariance of standard quantitative model based on texture and objectdatabases, and to compare with the popular computer vision method, andfinally to obtain valuable conclusions. The major research work and criticalconclusions includes:First, from the view of biological vision, we introduce the ventralpathway, receptive field and its features, and analyze the invariance and selectivity of the visual cortex, and clarify the biological basis of thestandard model.Second, from the perspective of computer vision, we respectivelyintroduce the biological visual hierarchical model, the local binary patternfeature, SIFT feature matching algorithm. We focus on analyzing theperformance of the model and its invariance and selectivity. Meanwhile, weintroduce the evaluation method to this paper.Third, evaluation and analysis of object and texture classification.Firstly, we evaluate the size of training sets, parameter and the run-time, thenanalyze the invariance and selectivity of texture databases, object databasesand high-resolution remote sensing databases. Then, compare the invariancewith LBP feature and SIFT feature to obtain a conclusion.①A great deal of experiments show that the biological visual featuresextracted by the standard model have superior rotational, scale and affineinvariance to textured images.②Comparing with the popular computer visual method, e.g. LBP andSIFT features, it’s found that: the rotation invariance of SMFs is far betterthan LBP feature, slightly better than SIFT feature; the scale invariance ofSMFs feature is better than LBP feature. For SIFT feature, when the scalelevel is large, the scale feature of standard model is better than SIFT feature.When the scale level is small, the scale invariance of SIFT feature is better; the affine invariance of SMFs feature is better than LBP and SIFT algorithm.③Based on this, the paper also carry out evaluation of invariance andselectivity of each level through the ventral pathway. The experimentalresults show that, the trends of invariance and selectivity of SMFs feature onthe ventral pathway is not consistent with the trends of invariance andselectivity of biological vision system, which indicates that the SMFs modelcan not reflect the trend of the invariance and selectivity on the ventralpathway.