電子情報通信学会総合大会講演要旨
D-12-41
分類器の予測確率情報を用いた葉画像分類の提案
○Siang Thye Hang・立間淳司・青野雅樹(豊橋技科大)
In this paper, a method utilising probability estimation by LibSVM in leaf scan image classification is proposed. Leaf scan images in the PlantCLEF dataset are used. In the pre-processing stage, besides binarisation with Otsu method, bilateral filtering and connected component extraction are applied as an improved masking strategy. With the stated pre-processing strategy, evaluation results of conventional features on the leaf scan images showed that HOG performed the best (63%), followed by densely SIFT (DSIFT) encoded with Fisher Vector (FV) (62%). Concatenation of both HOG and FV-DSIFT resulted in 67% accuracy; while with multiplication of estimated probability, classification accuracy is further improved to 70% in multiple datasets.