Machine Learning - Special issue on applications of machine learning and the knowledge discovery process
An introduction to ROC analysis
Pattern Recognition Letters - Special issue: ROC analysis in pattern recognition
A systematic analysis of performance measures for classification tasks
Information Processing and Management: an International Journal
| Hi-index | 0.00 |
Knots on digital images of 51 full veneer sheets, obtained from nine peeler blocks crosscut from two 35-foot (10.7m) long logs and one 18-foot (5.5m) log from a single Douglas-fir tree, were detected using a two-phase algorithm. The algorithm was developed using one image, the Development Sheet, refined on five other images, the Training Sheets, and then applied to all remaining sheets. In phase one, global thresholding was used to segment the image through a series of morphological operations to isolate regions likely to contain knots. In phase two, adaptive thresholding was applied to grey scale and red component segmented images to improve the accuracy of the segmented knot. Overall performance, judged in terms of confusion matrix performance metrics, was better for the red component images. Red component recall (true positive) rate was 1.00, 0.99, and 0.96 for the Development, Training, and complete sets, respectively. For the grey scale images, recall rates were 0.96 for all sets. Red component accuracy was 0.76, 0.92, 0.73 (Development, Training, and complete) and those for the grey scale images were 0.71, 0.85, and 0.69, respectively. Red component precision also exceeded that of the grey scale (0.75, 0.93, 0.73 compared to 0.72, 0.88, 0.70). A greater percentage of knots (78%) segmented from red component images were correctly sized, while 16% had more pixels than required and 6% had fewer pixels. Comparative figures for the grey scale images were 57% correctly sized, 2% with more pixels, and 42% with less pixels. Based on our results, we will adopt the red component image for continuing work with digital veneer images from a sample of Douglas-fir trees selected on the basis of acoustic velocity measures. Together with acoustic measurements of the veneer sheets, we are investigating the extent that the number, size, and spatial arrangement of knots influences the average stiffness of veneer sheets, with a view to determining if a relationship exists between the average stiffness of veneer sheets in a peeler block, stiffness of the log, and stiffness of the parent tree from a range of silvicultural treatments.