http://chineseinput.net/에서 pinyin(병음)방식으로 중국어를 변환할 수 있습니다.
변환된 중국어를 복사하여 사용하시면 됩니다.
< 구두-B-02 > A Deep Learning Approach to Wood Anatomy
( Kayoko Kobayashi ),( Sung-wook Hwang ),( Junji Sugiyama ) 한국목재공학회 2018 한국목재공학회 학술발표논문집 Vol.2018 No.1
Machine learning has been applied in the various fields; especially image recognition has been developing rapidly. We reported that such a technique is applicable for non-destructive wood identification of cultural important things<sup>1)</sup>, and also showed the potential to give us a novel aspect of wood anatomy<sup>2)</sup>. In these previous works, we used the ‘traditional’ machine learning method, such as gray level co-occurrence matrix and scale invariant feature transform. On the other hand, in the present study we present a deep learning approach, which has been attracted much attention for the last decades. Since the deep learning requires a large amount of data, we prepared the optical micrographs of transverse sections for 113 hardwood species, which belong to Fagaceae, Magnoliaceae, Lauraceae, etc. The classification accuracy at species level reached over 95% when using the images with the resolution of 3 μm/pixel and the VGG16 pre-trained model.
Characterization of crystalline linear (1→3)-α-<small>D</small>-glucan synthesized <i>in vitro</i>
Kobayashi, Kayoko,Hasegawa, Takuto,Kusumi, Ryosuke,Kimura, Satoshi,Yoshida, Makoto,Sugiyama, Junji,Wada, Masahisa Applied Science Publishers 2017 Carbohydrate polymers Vol.177 No.-
<P><B>Abstract</B></P> <P>We investigated the crystal structure and molecular arrangement of the linear (1→3)-α-<SMALL>D</SMALL>-glucan synthesized by glucosyltransferase GtfJ cloned from <I>Streptococcus salivarius</I> using sucrose as a substrate. The synthetic products had two morphologies: wavy fibril-like crystals as major and thin lamellae as minor products. Their structures were analyzed using electron microdiffraction, synchrotron X-ray powder diffraction, and solid-state <SUP>13</SUP>C NMR spectroscopy. The fibrils and lamellae had the same allomorphic form but different molecular arrangements. The wet crystals were in a hydrated form, which converted into an anhydrous form with a significant decrease in crystallinity on drying. The hydrated and anhydrous forms had an extended-chain conformation with 2/1 helix, and the hydrated form was estimated to contain one water molecule per glucose residue. The long glucan chains were folded in the fibril crystals, while the short, extended chains were arranged perpendicular to the base plane of the lamellae.</P> <P><B>Highlights</B></P> <P> <UL> <LI> (1→3)-α-<SMALL>D</SMALL>-Glucan was synthesized by recombinant glucosyltransferase sing sucrose. </LI> <LI> The synthetic products had two morphologies: wavy fibril and thin lamella. </LI> <LI> Both products had the same crystal structures but different molecular arrangements. </LI> <LI> Molecular chains were folded in the fibril but extended in the lamellar crystals. </LI> <LI> The hydrated form was converted into an anhydrous form by drying. </LI> </UL> </P>
( Takeshi Nakajima ),( Kayoko Kobayashi ),( Junji Sugiyama ) 한국목재공학회 2019 한국목재공학회 학술발표논문집 Vol.2019 No.1
Tree-ring analysis is an important field of science, including dendrochronology, dendroclimatology and modeling the tree growth environmental response system. In most cases the analyses have been conducted using one parameter from one tree-ring, e.g. ring-width, density, ratio of radioisotope, and so on. The information within a ring, however, has been less studied and many more things to be explored such as seasonal response in the shorter time scales. From another point of view, many species of softwood are often used into tree-ring analyses but our previous work revealed that simple CNN models did not work well in identification of softwood images where the morphology is rather regular or periodic. Therefore, substantial improvement in either feature extraction or the design of neural network was needed. In this study, therefore, we applied wavelet transform into deep-learning technique in order to extract information of tree growth environmental response in sub-seasonal time scales from softwood images.
( Sung-wook Hwang ),( Kayoko Kobayashi ),( Junji Sugiyama ) 한국목재공학회 2019 한국목재공학회 학술발표논문집 Vol.2019 No.1
A bag-of-features (BOF) model was implemented to further investigate the Lauraceae family through computer vision techniques. The key function of the BOF is to represent an image as a histogram using a codebook (or visual words) generated by clustering of detected features from images. Local features extracted by the scale-invariant feature transform (SIFT) algorithm were used to generate a codebook. In our previous study, it was confirmed that the SIFT keypoints have a high discriminant power in wood classification. In the proposed model, 1019 cross-sectional optical micrographs of 9 species across 6 genera from the Lauraceae family are tested. The visual words can be further analyzed by mapping them to each image to visualize the corresponding anatomical features. From such analysis, the computer vision technique can classify the aggregation of different combinations of wood cells such as vessels, wood fibers, rays, and axial parenchyma cells. The term frequency-inverse document frequency weights reveals that cell corner-based features are more species specific than cell lumen-based features. The codebook-based wood recognition model allowed us to approach the classification problem of wood based on the domain knowledge of wood anatomy.
< 구두-B-04 > Species-specific features in bag-of-features model
( Sung-wook Hwang ),( Kayoko Kobayashi ),( Junji Sugiyama ) 한국목재공학회 2018 한국목재공학회 학술발표논문집 Vol.2018 No.1
Bag-of-features (BOF) model is one of the computer vision-based method to classify image, and this model was derived from the bag-of-words model developed for automatic classification of documents. The BOF is simple and powerful, so it widely used for image classification. In this study, we tried to identify Lauraceae image database. The database including 1658 optical micrographs form 11 genera with 39 species. Lauraceae is known as a family that is difficult to identify because of its vast variety of species and similar features. It is the reason why we want to identify this family by the computer vison-based method. The image features were extracted using the scale-invariant feature transform (SIFT) algorithm, and the identification was performed by the support vector machine (SVM). To find the species-specific feature, we used tf-idf (term frequency - inverse document frequency) score weight. The score of common features are reduced, while score of rare, unique, and important features are increased. From the tf-idf score, therefore, we can find the species-specific features. The identification accuracy of the BOF was excellent at 98.2%. The accuracy is slightly higher than the 95.4% accuracy of our previous study that used only the SIFT algorithm and SVM without using the BOF. The tf-idf score showed important features for each species, and the features varied from species to species. Furthermore, the species-specific features differed by species in the same anatomical features and even showed area-specific characteristics within the same anatomical features of the same species. From the BOF model with the tf-idf score, we were able to better understand what the computer looks from the image.