Abstract Deep learning based computer vision is regarded as a promising method to achieve automatic wood defects detection. However, due to the complexity and diversity of wood natural defects, the recognition accuracy is always not satisfactory to practical application. It is believed that it is the lack of specific features of wood defects that causes the lower accuracy of deep learning model. In this study, traditional activation function of CNN model is substituted by non-monotonic functions in order to achieve a better approximation ability which is though will help model to capture the key features of wood defects. The principle of Kolmogorov-Arnold Networks(KAN) is also adopted to parameterize and discretize the activation function to make it learnable and facilitate model training. Our findings indicate that some non-monotonic functions do outperform traditional activation function, ReLU. Parameterization of activation function can further improve the accuracy. By utilizing Taylor series, activation functions are discretized, and it is found that increasing the order of Taylor series will also enhance the accuracy. Specifically, new type of activation function highlights key features and can better approximate the relationships among variables, making a contribution to wood defects recognition. Moreover, by quantifying highlights of the feature maps achieved by new activation function, the specific features associated with wood structures are obtained. The findings of this study not only provide us with the guidance on improving accuracy of wood defects recognition but also make deep learning model explainable which improves its reliability.
Wang et al. (Mon,) studied this question.