<p>Deep learning based computer vision is regarded as a promising method to achieve automatic wood defect detection. However, due to the complexity and diversity of wood natural defects, the recognition accuracy fails to meet practical application. It is believed that the lack of specific features of wood defects causes the lower accuracy of deep learning model. In this study, the traditional activation function of CNN model is substituted by the non-monotonic functions in order to achieve a better approximation ability which is expected to 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 its training. Our findings indicate that some non-monotonic functions 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 to a certain extent will also enhance the accuracy. Specifically, the new type of activation function highlights key features and can better approximate the relationships among variables, contributing to wood defect recognition. Moreover, by quantifying highlighted regions 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 guidance on improving accuracy of wood defect recognition but also make deep learning models explainable which improves their reliability.</p>