Comparative study among three strategies of incorporating spatial structures to ordinal image regression

Abstract

Images usually have specific spatial structures, and related researches have shown that these structures can contribute to the establishment of more effective classification algorithms for images. So far though there have been many solutions of making use of such spatial structures separately proposed, little attention has been paid to their systematic summary, let their comparative study alone. On the other hand, we find that the existing image-oriented ordinal regression (OR) methods do not utilize such structure information, which motivates us to compensate a comparative study through embedding such spatial structure into ORs. Towards the end, in this paper, we (1) through a summary, find three typical strategies of using image prior spatial information, i.e., structure-embedded Euclidean distance strategy, structure-regularized modeling strategy for classifier learning, and direct manipulation strategy on images without vectorization for image; more importantly, (2) apply these strategies to establish corresponding ORs for classifying data with ordinal characteristic, conduct comprehensive comparisons and give analysis on them under three evaluation criteria. Experimental results on typical ordinal image datasets JAFFE, UMIST and FG-NET show that the latter two strategies can, on the whole, achieve distinct gain in OR performance and while the first one cannot necessarily as expected, which is due to whether the spatial information is directly embedded into the objective function involved or not.