Effect of moisture content variation on CT image classification to identify internal defects of sugar maple logs.
Sugar maple logs (Acer saccharum Marsh.) were scanned with a medical x-ray scanner in order to evaluate the relationship between the moisture content (MC) and the performance of a parametric supervised classification procedure to detect internal defects. A first group of logs coming from five trees harvested at the end of May was scanned after 0, 2, 6, 10, and 14 weeks of storage. A second group of logs coming from five trees harvested at the end of August was scanned only once, immediately after felling. The CT images selected from each log were classified and filtered in order to identify and separate the sapwood from other internal characteristics. The sapwood and overall accuracies were calculated from a confusion matrix that was determined for each image. For logs of group 1, the storage time resulted in lower MC. No significant differences in MC were observed between groups 1 and 2 for the freshly cut condition. A quadratic relationship was established between MC and detection accuracy from a regression analysis. The coefficients of determination ([r.sup.2]) were 0.58 and 0.42 for sapwood and overall detection accuracy, respectively. The maximum sapwood detection accuracy was 0.93, which was reached at 62 percent MC. Between 44 and 67 percent MC, sapwood detection accuracy exceeded 0.80. These results indicate that the MC variations in sugar maple logs have an important impact on defect-detection capability of this method.
The identification of internal defects in logs before sawing should allow the sawmill to maximize value recovery. As a rule, the commercial value of hardwood lumber is inversely related to the quantity and sizes of defects that are present (Sarigul et al. 2003). To optimize lumber value recovery, however, producers are increasingly interested in nondestructive internal inspections of logs. Computed tomography (CT) has shown potential for detecting internal defects and outer shape irregularities on sawlogs. The ability of this method to quantify spatially the density variations in solid wood, the presence of internal defects and the variation of moisture content (MC) in wood specimens explains the increasing interest of wood science and technology researchers (Bucur 2003).
A variety of methods are being used to detect and identify defects in CT images (Guddanti and Chang 1998; Bhandarkar et al. 1999; Schmoldt et al. 1995, 1998, 2000; Nordmark 2002, 2003). However, the problem is difficult to solve because of the inherent variability, complex structure of wood, and MC variations, so that complete success has not yet been achieved. Knowledge of wood density usually makes it possible to differentiate between the characteristics present in logs, such as knots, bark, decay, sapwood, heartwood, voids, etc. However, the capability of differentiating among these characteristics in CT scan log images is also influenced by MC. Schmoldt et al. (1995) indicated that, as CT number (grey level) is directly related to density, CT image grey levels vary with different species and MC. As a consequence, a log that is freshly cut will produce different CT number values than one that has had time to dry. Oja and Temnerud (1999) reported that, for green logs of Norway spruce, the grey level of normal wood differs very much between heartwood and sapwood because of the difference in MC. Rojas et al. (2005) observed a similar situation in CT images of sugar maple logs. Sapwood generally showed two regions: an inner region, with a slightly clearer shade near the colored heartwood, and an outer region, which was darker. This contrast in grey level was explained mainly by the lower MC levels in the outer area due to drying in storage in the absence of bark.
Rojas et al. (2006) have recently proposed a parametric supervised classification procedure for identifying and separating sapwood from colored heartwood, knots, rot, splits, and bark in sugar maple logs. This classification procedure uses only grey level information for labeling each pixel of the CT image as either sapwood or any particular type of defect. Sapwood was identified and separated from any defect with an accuracy of 97 percent for group 1 (freshly cut logs) and 82 percent for group 2 (unknown date of harvesting), respectively. The study also indicated that sapwood accuracy values were very sensitive to MC variations. The objective of the present work was therefore to establish the relationship between MC levels in sugar maple logs and the accuracy achieved in the detection of internal defects from these logs using a parametric supervised classification method.
Materials and methods
Logs selection scanning
Ten sugar maple (Acer saccharum Marsh.) trees were chosen from one stand in the Duchesnay forest, 45 km North-East of Quebec City, Quebec. Five trees were harvested at the end of May (group 1) and the other five at the end of August (group 2). Each tree was bucked into three 5-foot long logs. Log height class was recorded, log height values of 1, 2 and 3 being assigned to the butt, mid and top logs. Disks 25 mm in thickness were immediately cut from both ends of each log and stored in plastic bags for initial MC measurement. Three other 25-mm thick disks were then selected and marked on the surface of each log in order to evaluate the presence of internal defects. In total, 45 disks were evaluated for each group of trees. The logs of group 1 were weighed and scanned (only the selected and marked disks) after 0, 2, 6, 10 and 14 weeks. During this period the logs remained stored outdoors and protected against the rain. The green mass of each log was measured to the nearest 10 g. The group 2 logs were scanned only once. Scanning was performed with a Siemens Somaton x-ray CT scanner (Rojas et al. 2006). Then, five consecutive CT images, equivalent to the thickness of each disk selected, were used to study the presence of rot, knots, colored heartwood and sapwood.
Initial MC at felling time was estimated from two disks cut from the ends of each log. The green mass of each disk was measured to the nearest 0.001 g. All disks were then oven-dried at 103[degrees]C for 36 hours. After cooling to room temperature, ovendry mass was measured. The MC of each disk was expressed as a percentage of ovendry mass. The two disks' MC were averaged and used as an estimation of the whole log's MC. Based on this log MC estimation and green mass, a log dry mass was also estimated. The MC of each log at each scanning time was estimated from the green mass measured before each scanning operation and the log dry mass estimate.
Classification of CT images
All CT images selected from both groups of logs were segmented using the parametric supervised classification method suggested by Rojas et al. (2006). This method of internal defect detection is based on a priori knowledge of the spectral signature (gray level) of each internal defect. The sapwood and overall detection accuracy values were obtained from a confusion matrix determined for each CT images. The classification procedure and quantitative evaluation were performed using the PCI software (PCI User Guide 1997).
The results were evaluated using the SAS software. Data obtained for MC were analyzed using a Mixed ANOVA model. Since we were mainly interested in validating the CT image detection method proposed by Rojas et al. (2006) to identify and segment the sapwood, a regression analysis was conducted between MC and accuracy values obtained over the weeks of log drying.
Results and discussion
A typical series of CT images obtained for the same cross section from group 1 logs after 0, 2, 6, 10 and 14 weeks of felling is shown in Figure 1. Figure 1a clearly shows internal log characteristics such as knots, colored heartwood, splits, sapwood, and bark. Figures 1b, 1c, 1d, and 1e show variations in the signature while log drying is taking place. The dark areas observed in these consecutive CT images generally correspond to areas of lower MC. When Figures 1a and 1e are compared visually, it can be observed that the sapwood zone presents more variation in the spectral signature than the colored heartwood zone. For the colored heartwood, the principal changes occurred in the central zone, and those changes could be explained by the presence of an internal split. Less variation in the spectral signature was observed in the knots and bark. Spectral signature variation can also be observed in the bark surroundings. This is explained by MC variation due to bark separation. Overall, these results confirm the information reported by Rojas et al. (2005). In spite of the variation in spectral signatures, it appears possible to identify the sapwood limit and also some of the other internal characteristics.
[FIGURE 1 OMITTED]
The MC values determined for logs from group 1 trees and group 2 are presented in Table 1. For group 1, the results show values obtained after 0, 2, 6, 10 and 14 weeks of storage. At felling time (0 wk), the trees in this group did not exhibit high MC variations. The minimum and maximum mean values were of 54.8 percent and 62.8 percent MC and were obtained from trees 2 and 3, respectively. Contrary to expectations, similar MC were observed for group 2, in which trees 3 and 5 exhibited the minimum (54.3%) and maximum (66.2%) mean values respectively. We had expected higher MCs in trees from group 1 (cut in May). Clark and Gibbs (1957) reported maximum MC values of 80 percent between April and May for three different species. The lack of a difference in MC between the two groups of trees could not be attributed to any specific cause. We looked at the meteorological pattern for the location and time of harvest but it did not explain the high MCs observed with the August harvest.
For group 1 logs, the main decrease in MC was generally observed between the first 2 weeks, with a mean value of 6.9 percent MC percentage points. By contrast, a relatively small decrease of 2.2 percentage points was observed between the tenth and fourteenth weeks of storage. The mean total MC loss after 14 weeks of storage was determined to be 22 percentage points.
An analysis of variance on group 1 trees indicated that storage time was a significant source of MC variation (Table 2). It also indicated that MC differences among log height classes were not significant. MC differences within merchantable tree height were also not significant for trees of group 2. Finally, a comparison of MC between groups 1 and 2 after felling indicated no significant differences.
Relationship between MC and accuracy classification
The results of mean sapwood identification and overall defect detection accuracy values obtained for logs from groups 1 and 2 trees are shown in Table 3. In this table, each accuracy value represents the mean value obtained from three CT images. Sapwood accuracy was determined from the confusion matrix dividing the number of sapwood pixels correctly classified by the total number of this class in the reference data. Overall detection accuracy was calculated by dividing the total of pixels correctly classified as sapwood, colored heartwood, knots, rot, split, and bark by the total number of pixels in the confusion matrix (Rojas et al. 2006).
For both groups of trees, the accuracy values obtained for sapwood were higher than the overall accuracy (Table 3). This is explained by the fact that the overall accuracy is the result obtained for the detection of six classes (sapwood, colored heartwood, knots, rot, splits, and bark), and each class normally presents an error label. These errors are mainly due to an overlap in the spectral signature observed between colored heartwood and knots, and occasionally between colored heartwood and bark (Rojas et al. 2005 and 2006). The classification procedure operates on a pixel-by-pixel basis, and the main source of classification error is associated with the omission error--for instance, pixels corresponding to colored heartwood but omitted from this class. We also observed an overlap between the spectral signatures of sapwood and rot (Rojas et al. 2005). This overlap appears when the sapwood spectral signature decreases, which is related to any MC loss (Fig. 1). However, an overlap can also occur between sapwood and colored heartwood, mainly when the colored heartwood zone is very heterogeneous. In this case it could be explained by either the presence of central rot or by MC variations. Such heterogeneity will clearly permit an overlap with the spectral signature of sapwood and a decrease in detection accuracy values. This result agrees with those reported by Schmoldt et al. (1995), Oja and Temnerud (1999), and Rojas et al. (2006).
The results presented in Table 3 show that accuracy is influenced by MC variations in logs. Detection accuracy values decrease as log storage time increases. This behavior shows the limits of the detection method proposed by Rojas et al. (2006) with respect to log MC variations.
A regression analysis was carried out between MCs and detection accuracy values (sapwood and overall). Initially, the empirical Logit transformation was used on data in order to obtain predicted values between 0 and 1. The MC was used as the independent variable and detection accuracy value (AV) as the predicted variable in a model of the form:
AV = exp([B.sub.0] + [B.sub.1]MC + [B.sub.2]M[C.sup.2])/1 + exp([B.sub.0] + [B.sub.1]MC + [B.sub.2]M[C.sup.2]) 
The results of the regression analysis for all scans of trees are shown in Table 4. The coefficients of determination ([r.sup.2]) for sapwood and overall detection accuracy were 0.58 and 0.42 respectively. At the same time, coefficients of variation (CV) of 12 percent and 9 percent respectively were obtained for sapwood and overall detection accuracy. Although the [r.sup.2] values appear relatively low, the CV values are low enough to allow for the use of the model for predictive purposes.
The relationships between MC and accuracy (sapwood and overall detection) for all logs are shown in Figures 2 and 3, respectively. Since a MC comparison after felling indicated no significant differences between groups 1 and 2, they were pooled in the regression analysis. In both Figures, it can be observed that a quadratic relationship provides the best fit for the relationship between MC and accuracy.
[FIGURES 2-3 OMITTED]
Figures 2 and 3 show minimum and maximum MC values of 32 percent and 67 percent. In the case of sapwood detection (Fig. 2), the accuracy estimate values for these MC were 0.58 and 0.91 respectively. We determined the value of maximum accuracy by calculating the maximum value of the quadratic equation. The maximum sapwood detection accuracy of 0.93 was found to be at 62 percent MC. Figure 2 indicates that sapwood accuracy values higher than 0.8 can be obtained from logs having MCs between 44 percent and 67 percent. The same analysis for the overall detection accuracy (Fig. 3) indicated that the maximum overall detection accuracy of 0.81 was reached at 53 percent MC, while the accuracy values for 32 percent and 67 percent of MC were of 0.53 and 0.64 respectively.
Training set determination from CT images is the main step of the supervised classification procedure. The spectral signature for each class obtained from this training set is used to obtain decision functions by the classification algorithm. In the case of wood, the information also depends on log MC. In the present study, we applied the classification procedure developed previously by Rojas et al. (2006) to a group of CT images from a new set of logs. The logs where the training set had been determined in the previous study were in the order of 60 to 70 percent MC. This explains why we obtained maximum detection accuracy, toward the highest MC values in the second trial, at 53 percent and 62 percent respectively for overall and sapwood detection.
These results show that the detection potential of sapwood and of other internal defects using a parametric supervised classification procedure is affected by MC variations in sugar maple logs. This situation can be explained by the physical principle of the CT scanner. The information obtained from a CT image is associated to the density of the object. It is known that green density of logs is affected by MC. It is hence established that the variation of this parameter represents the main limitation of this method.
In the future, however, it would be advisable to generate a new training set from a larger database of sugar maple logs and then evaluate this classification procedure using this training set. Broadening the database would make the detection method more robust. This modification would improve the performance of this procedure and at the same time increase the feasibility of using it in the future.
This paper presents a validation of a parametric classification procedure previously proposed to identify and separate sapwood from colored heartwood, knots, rot, splits, and bark in sugar maple logs. We reported that sapwood detection accuracy was very sensitive to MC variations. We therefore decided to explore the effects of MC variations in logs after several drying periods. No significant differences in MC were observed between groups 1 and 2 for 0 week of storage time. This was unexpected since our aim was to provide contrasts between the MCs of logs felled in spring and in late summer. The study determined that variations in spectral signature of CT images for internal log defects can be associated with the loss of MC during storage.
A quantitative evaluation indicated that drying during storage normally influenced the sapwood and overall detection values. A regression analysis from group 1 led to a quadratic relationship between MC and detection accuracy (sapwood and overall). This regression model generally indicated that sapwood can be detected with an accuracy level of at least 0.80 when the log MC is between 44 percent and 67 percent. Minimum and maximum sapwood accuracy value of 0.58 and 0.92 were obtained for MC of 32 percent and 62 percent respectively. For overall defect detection accuracy, the results indicated that the maximum prediction value was 0.81 for MC of 53 percent. The accuracy values obtained for 32 percent and 67 percent MC were 0.53 and 0.64, respectively.
This study shows that the parametric classification procedure previously proposed presents significant potential for detecting sapwood in sugar maple logs. However, its performance is limited to a moisture range of 44 percent to 67 percent. In future, it would be advisable to generate a new training set from lower MC sugar maple logs in order to see how sapwood prediction accuracy could be enhanced for lower MCs. Broadening the data set would make the detection method more robust. This modification would improve the performance of this procedure and the feasibility of using it. The question of MC determination prior to scanning would however remain.
Bhandarkar, S.M., T.D. Faust, and M. Tang. 1999. CATALOG: A system for detection and rendering of internal log defects using computer tomography. Mach. Vis. Appl. 1999(3):171-190.
Bucur, V. 2003. Nondestructive characterization and imaging of wood. Springer-Verlag Berlin Heidelberg. 354 pp.
Clark, J. and R.D. Gibbs. 1957. Studies in tree physiology. IV. Further investigations of seasonal changes in moisture content of certain Canadian forest trees. Can. J. Bot. 35:219-253.
Guddanti, S. and S.J. Chang. 1998. Replicating sawmill sawing with TOPSAW using CT images of a full-length hardwood log. Forest Prod. J. 48(1):72-75.
Nordmark, U. 2002. Knot identification from CT images of young Pinus sylvestris sawlogs using artificial neural networks. Scand. J. Forest Res. 17:72-78.
--. 2003. Models of knot and log geometry of young Pinus svlvestris sawlogs extracted from Computed Tomographic Images. Scand. J. Forest Res. 18:168-175.
Oja, J. and E. Temnerud. 1999. The appearance of resin pocket in CT-images of Norway spruce (Picea abies (L.) Karst.). Holz als Roh- und Werkstoff 57:400-406.
PCI User Guide. 1997. Volume I. Version 6.1.50 West Wilmot street, Richmond Hill, Ontario, Canada.
Rojas, G., R.E. Hernandez, A. Condal, D. Verret, and R. Beauregard. 2005. Exploration of the physical properties of internal characteristics of sugar maple logs and relationships with CT images. Wood and Fiber Sci. 37(4):591-603.
--, A. Condal, R. Beauregard, D. Verret, and R.E. Hernandez. 2006. Identification of internal defects of sugar maple logs from CT Images using supervised classification methods. Holz als Roh-und Werkstoff. (published online
Sarigul, E., A.L. Abbott, and D.L. Schmoldt. 2003. Rule-driven defect detection in CT images of hardwood logs. Comput. Electron. Agric. 41:101-119.
Schmoldt, D.L., P. Li, and A.L. Abbott. 1995. Log defect recognition using CT-images and neural net classifiers. 2nd Inter. Workshop/ Seminar on Scanning Tech. and Image Processing on Wood. Skelleftea, Sweden, Aug 14-16. pp 77-97.
--, J. He, and A.L. Abbott. 1998. A comparison of several artificial neural network classifiers for CT images of hardwood logs. Pages 34-43. Machine vision applications in industrial inspection VI. Vol. 3306. The Inter. Soc. for Optical Engineering (SPIE). www.spie.org/
--, --, and --. 2000. Automated labeling of log features in CT imagery of multiple hardwood species. Wood and Fiber Sci. 32(3):287-300.
Gerson Rojas Robert Beauregard * Roger E. Hernandez * Daniel Verret Alfonso Condal
The authors are, respectively, Former PhD Candidate, Centre de recherche sur le bois (CRB) Departement des sciences du bois et de la foret Universite Laval, Quebec, Quebec, Canada G1K 7P4 and Assistant Professor, Departamento de Ing. en Maderas, Universidad del Bio-Bio, Av. Collao 1202, Casilla 5-C, Concepcion, Chile (email@example.com); Associate Professor, Professor, Centre de recherche sur le bois (CRB), Centre de recherche sur les technologies de l'organisation reseau (CENTOR), Departement des sciences du bois et de la forat, Universite Laval, Quebec, QC, Canada G1K 7P4 (Robert.Beauregard@sbf.ulaval.ca; Roger.Hernandez@sbf.ulaval.ca); Researcher (Currently President of ForwardSim, Inc.) Forintek Canada Corp., 319 rue Franquet, Quebec, QC, G1P 4R4 (firstname.lastname@example.org); and Professor, Departement des sciences geomatiques, Universite Laval, Quebec, QC, Canada G 1K 7P4 (Alfonso.Condal@scg.ulaval.ca). This paper was received for publication in December 2005. Article No. 10142.
* Forest Products Society Member.
Table 1.--Summary of MC values for logs of groups 1 and 2 of sugar maple trees. Moisture content at time after felling Group 1 Group 2 0 2 6 10 14 0 Tree Log weeks weeks weeks weeks weeks weeks (%) 1 1 55.9 49.0 42.8 38.0 36.0 59.0 2 55.9 51.7 46.1 41.1 39.0 60.0 3 55.7 50.9 44.8 39.1 36.6 56.7 2 1 54.6 47.5 42.3 37.1 36.3 58.0 2 53.6 46.2 41.3 36.6 34.2 58.1 3 56.4 48.5 43.4 39.2 36.8 59.0 3 1 65.4 56.1 49.0 42.0 38.5 53.5 2 63.1 55.7 49.6 43.0 40.1 53.8 3 60.1 53.5 46.0 40.2 37.3 54.6 4 1 57.6 51.3 45.6 39.9 37.5 60.7 2 58.3 52.5 46.5 40.7 36.9 62.7 3 59.4 52.7 46.6 42.1 39.6 62.7 5 1 57.1 47.9 42.5 36.5 33.9 64.2 2 56.4 47.9 41.5 35.6 31.9 66.7 3 58.2 51.3 44.7 36.7 34.2 67.0 Table 2.--Mixed ANOVA model of MC values of logs obtained from sugar maple trees of group 1. Source of variation DF F value Pr > F Log height position 2 0.21 0.8120 Time of storage 4 720.69 <0.0001 Log height position x time of storage 8 0.31 0.9580 Table 3.--Summary of mean sapwood detection and overall defect detection accuracy values for logs of groups 1 and 2 of sugar maple trees. Group 1 0 weeks 2 weeks 6 weeks Tree Log SA (a) SO (b) SA (a) SO (b) SA (a) SO (b) (%) 1 1 94.3 84.8 98.3 90.5 84.1 78.1 2 92.8 80.5 94.5 85.7 95.2 83.3 3 95.6 84.3 96.8 86.2 70.2 63.3 2 1 90.5 71.0 96.0 80.0 91.6 79.5 2 90.4 73.8 92.7 83.8 83.8 74.8 3 90.2 81.4 93.6 86.2 87.2 78.1 3 1 84.0 69.5 96.2 83.7 77.0 71.4 2 93.5 74.3 93.1 83.2 88.0 80.6 3 95.9 77.6 92.0 82.1 81.3 73.5 4 1 69.9 63.8 91.6 83.3 91.4 80.5 2 77.6 62.9 84.3 76.7 87.5 78.1 3 78.0 61.5 79.2 71.9 84.4 78.1 5 1 84.2 73.3 85.9 72.2 80.0 69.0 2 77.7 68.3 84.5 81.5 85.5 73.3 3 80.4 73.5 74.0 72.8 85.1 80.0 Group 1 Group 2 10 weeks 14 weeks 0 weeks Tree SA (a) SO (b) SA (a) SO (b) SA (a) SO (b) (%) 1 71.8 63.8 61.1 51.4 78.1 60.0 75.2 69.0 67.9 60.0 84.7 65.3 60.6 51.9 46.8 41.4 79.2 69.5 2 75.7 68.6 67.2 65.7 97.4 77.1 82.3 74.3 72.8 67.1 93.8 71.9 81.6 76.7 61.8 60.0 91.3 72.9 3 83.2 78.1 63.6 60.7 51.7 41.9 84.4 79.5 64.3 61.8 52.1 48.6 78.5 75.1 62.5 59.8 61.7 59.5 4 81.8 76.7 64.3 58.1 99.4 87.1 70.9 61.9 59.3 54.3 99.3 88.1 81.2 72.9 76.3 69.0 96.3 84.8 5 76.7 61.5 59.4 55.2 98.0 81.4 67.9 63.8 58.9 55.5 98.8 88.1 68.1 64.2 60.0 53.8 94.9 82.4 (a) SA--sapwood accuracy. (b) SO = overall accuracy. Table 4.--Regression analysis for sapwood and overall detection accuracy for all scans. Regression model Estimate Error df t value Sapwood detection Intercept -4.976 2.923 8 -1.7 MC 0.207 0.122 76 1.7 MC x MC -0.001 0.001 76 -1.7 Overall detection Intercept -5.236 1.688 8 -3.1 MC 0.235 0.070 76 3.1 MC x MC -0.002 0.001 70 -2.9 Regression model Pr > t [r.sup.2] C.V. Sapwood detection 0.58 12.1% Intercept 0.1270 MC 0.0933 MC x MC 0.2689 Overall detection 0.42 9.0% Intercept 0.0146 MC 0.0013 MC x MC 0.0046
|Printer friendly Cite/link Email Feedback|
|Author:||Rojas, Gerson; Beauregard, Robert; Hernandez, Roger E.; Verret, Daniel; Condal, Alfonso|
|Publication:||Forest Products Journal|
|Date:||Apr 1, 2007|
|Previous Article:||Evaluating loss of machinability in spalted sugar maple.|
|Next Article:||Effects of various fire-retardants on plate shear and five-point flexural shear properties of plywood.|