Image segmentation using fuzzy min-max neural networks for wood defect detection
In this work a colour image segmentation method for wood surface defect detection is presented. In an automated visual inspection system for wood boards, the image segmentation task aims to obtain a high defect detection rate with a low false positive rate, i.e., clear wood areas identified as defect regions. The proposed method is called FMMIS (Fuzzy Min-Max neural network for Image Segmentation). The FMMIS method grows boxes from a set of seed pixels, yielding the minimum bounded rectangle (MBR) for each defect present in the wood board image. The FMMIS method was applied to a set of 900 colour images of radiata pine boards, which included 10 defect categories. The FMMIS achieved a defect detection rate of 95 percent on the test set, with only 6 percent of false positives. The area recognition rate (ARR) criterion was computed, to measure the segmentation quality, using as a reference the manually placed MBR for each defect. The ARR achieved 94.4 percent on the test set. The results show significant improvements compared with previous work and that the computational load of FMMIS is suitable for real-time segmentation tasks.
Thank you for your questions:
A drawback of the FMMIS is that to keep the algorithm execution time low the number of input patterns (seeds) must not be high, this is because the algorithm is O(nh), where n is the number of seeds and h the number hyperboxes formed. But the user can define the maximum number of seeds required to satisfy the computational speed. Fot this case the maximum number of seeds were set to 100.
Future works are oriented to the application of the FMMIS to other applications, currently face detection. For the AVI system, a fourth layer to the FMMIS has been explored to perform classification of the segmented defects.
Thank you,
Gonzalo Ruz.
Have you used other methods that image segement to detect the defect. I remeber when I worked in farm people use sound and surfaced roughness.
You have specified that you give the coordinates of the seed pixels as inputs to the neural network. Is the colour intensity or some other property of the seed pixel is used in the evaluation process of the neural network? If so in which way?
Thank you for your question.
Log (wood) defect detection includes both internal and external analysis. Internal defect detection finds defects inside a log, significant internal defect detection research has been done using technology such as X-ray/CT, NMRI, ultrasound. External defect detection refers to locating defects on a log surface. Here you can find expensive devices such as laser scanning systems,which capture 3-dimensional topological information that contains log surface characteristics. Also more economical systems which use a simple video camera and apply image segmentation techiques on the 2D image of the wood.
More details related to different techniques for wood defect detection can be found in Journals such as FOREST PRODUCTS JOURNAL.
Our research has been oriented to external (surface) defect detection using a low-cost machine vision system composed of a colour video camera, a frame grabber, and a microcomputer.
Gonzalo Ruz.
Thank you Charles for your question.
When an input pattern (new seed) is presented, the hyperbox with the highest degree of membership is found and expanded to enclose the input pattern. The hyperbox expansion is accepted only if the region contained by the expanded hyperbox is similar in colour to the region enclosed by the hyperbox before the expansion. A fuzzy colour homogeneity criterion is defined to compare the colour similarity of two hyperboxes. This is based on the Z-function of the Euclidean distance of the mean colour intensities of the two hyperboxes, measured in the RGB space.
Gonzalo Ruz.
Hi Gonzalo,
Many congratulations on an excellent piece of work.
Keep it up!
Are there any drawback of this method or what is the future work for the current research?