...somewhat. We extracted, partitioned and described contours, histogrammed their geometric parameters and concatenated the histograms to form a single image vector with which we classi ed the plant images using a Linear Discriminant Analysis (LDA); that is, no segmentation or saliency selection was performed. Despite the obvious simplicity of the LDA classi cation we reached the middle of the ranking for sheet-as-background images. While contour-based feature extraction is presently still a lone-some strategy in comparison to the prevailing gradient-based matching techniques (e.g. SIFT), it may soon be a viable alternative - once we developed appropriate classi cation methodology to deal with the descriptors.
Whereas our classi cation was simple, our descriptor extraction is rather complex. We employed our method for curve partitioning and abstraction, which we previously introduced as part of an image classi cation study [Rasche, 2010]. Meanwhile the method has been improved and has also been applied in video indexing [Ionescu et al., 2012], image retrieval [Rasche and Vertan, 2010], gesture (posture) identi cation [Oprisescu et al., 2012] and in shape retrieval where we obtained the present benchmark [Rasche, 2013]. For reasons of time, we fell short of exploiting its full potential in the present plant classi cation task [Caputo et al., 2013,Goau et al., 2013]. Nevertheless, we obtain moderate results, with a simple classi cation of image vectors. We essentially pursued a structural description. Contours were extracted with the Canny algorithm [Canny, 1986] and then partitioned into (elementary) segments and geometrically described using a multi-resolution analysis. A contour is iterated with a xed window which measures the amplitude for its contour subsegment. The resulting local/global space is analyzed for consistent 'segments', which are identi ed as elementary segments, namely smooth arcs and straight(er) segments. These segments are then geometrically described by several parameters such as orientation, length, 'bendness' ( curvature), edginess, degree of alternating, etc.[Rasche, 2013]. The elementary segments were then grouped into a variety of pairs such as closure (2 curved segments facing each other as in '()' ), ribbons (2 parallel, aligned, straight segments), hyperbola (2 curved segments facing away as in ')(' ), etc. To avoid a combinatorial explosion, we selected symmetric pairs, which exhibited a high degree of bilaterality along either the symmetric axis or the mid axis. The pairs were then also geometrically parameterized (various distances between segments, their degree of the 'structural biases', etc.). For (elementary) segment and pair descriptors we also extracted (image) appearance parameters (contrast, standard deviation of pixels values along the contour, etc.). In total 100 to 150 parameters are determined for both segment and pair descriptors.
These segment and pair descriptors represent a 'structural alternative' to the prevailing gradient-based features (e.g. SIFT), which typically excel at textural representation. However we could not capitalize on the structural speci city yet, and therefore report only on a statistical discrimination. That is, for each image, the parameters across descriptors were histogrammed with 10 bins - and the individual descriptors were thus not actually exploited. The image vector had so a dimensionality of 1000 to 1500. 3
We applied the Principal Component Analysis to reduce the dimensionality to ca. 200-400 dimensions (depending on the classi cation task). A classical LDA was used, whereby we built a tree-like classi cation system. We rstly built a background classi er, which discriminated between natural and sheet background (99% on the training set). For natural-background images we then trained a type classi er discriminating between stem, leaf, entire, ower and fruit (ca. 70% on training set). Lastly, a (all-versus-all) species classi er was built to discriminate between the 250 species. For sheet-as-background images, we immediately applied a species classi er. For sheet-as-background images (species) classi cation reached the middle of the ranking (in comparison to the other approaches), and was also better than our performance for natural-sheet images. 4
Classifying plant images with a 'whole-image' classi cation approach as presented here will certainly not be a serious alternative, but we hope to improve on exploiting the individual contour parameters better in the future - as we did in other image collections - and not just histogram them. [Canny, 1986] Canny, J. (1986). A computational approach to edge-detection. IEEE
Transactions on Pattern Analysis and Machine Intelligence, 8(6):679{698. [Caputo et al., 2013] Caputo, B., Muller, H., Thomee, M., Villegas, R., Paredes, R., Zellhofer, D., Goeau, H., Joly, A., Bonnet, P., Martinez Gomez, J., Garcia Varea, J., and Cazorla, M. (2013). Imageclef 2013: the vision, the data and the open challenges.
In CLEF. CLEF. [Goau et al., 2013] Goau, H., Bonnet, P., Joly, A., Baki, V., Barthelemy, D., Boujemaa, N., and Molino, J. (2013). The imageclef 2013 plant identi cation task. In CLEF 2013 Working Notes, Valencia, Spain. [Ionescu et al., 2012] Ionescu, B., Seyerlehner, K., Rasche, C., Vertan, C., and Lambert, P. (2012). Content-based video description for automatic video genre categorization. In The 18th International Conference on MultiMedia Modeling, 4-6 January, Klagenfurt, Austria, Springer-Verlag LNCS - Lecture Notes in Computer Science, Eds. K. Schoe mann et al., number 7131, pages 51{62. [Oprisescu et al., 2012] Oprisescu, S., Rasche, C., and B., S. (2012). Automatic static hand gesture recognition using tof cameras. In 20th European Signal Processing Conference, Bucharest, RO. [Rasche, 2010] Rasche, C. (2010). An approach to the parameterization of structure for fast categorization. International Journal of Computer Vision, 87:337{356. [Rasche, 2013] Rasche, C. (2013). Curve partitioning and abstraction with the local/global space. IEEE Transaction on Image Processing., in revision. [Rasche and Vertan, 2010] Rasche, C. and Vertan, C. (2010). A novel structural-description approach for image retrieval. In CLEF (Notebook Papers/LABs/Workshops).