Research on Wood Defects Feature Imbalance Optimization and Recognition
Deep learning is a promising method to achieve automatic wood defects detection which is indispensable for wood production; however, such a technique faces challenges caused by poor generalization ability and low recognition accuracy on light defects. In this study, the problems are attributed to im...
Saved in:
| Main Author: | |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2025-01-01
|
| Series: | IEEE Access |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/10870126/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Deep learning is a promising method to achieve automatic wood defects detection which is indispensable for wood production; however, such a technique faces challenges caused by poor generalization ability and low recognition accuracy on light defects. In this study, the problems are attributed to imbalanced feature distribution of wood defects which are rich in diversity. To this end, two improvements on CNN classifier loss function are proposed to enhance the performance on wood defects recognition. Firstly, Cross Entropy loss is modified to allow the model to pay more attention to individual sample discrepancy during training. Secondly, a new loss, Var Loss is proposed to add to main loss function in order to decrease the variance of classification results, thus reducing the impact of sample diversity on model performance. Classical CNN classifiers are employed to distinguish three kinds of wood images: dead knot, live knot and ordinary grain. Results show that Modified Cross Entropy makes model more sensitive to hard samples regardless of the batchsize. Var Loss tends to decrease the fluctuation of prediction confidences, making model more robust in practical use. An overall accuracy increase of 6.6% is achieved, the accuracy on live knot has an increase of 17%, and the missing rate of defects is decreased by 15%. Besides, generalization ability test indicates that new methods allow the classifier to have comparable accuracies on five additional datasets. The proposed loss functions improve the model performance through optimizing the model training process, providing a new idea for deep learning application in wood defects detection. |
|---|---|
| ISSN: | 2169-3536 |