<p>In the field of signature recognition, different writers, writing styles, and even writing times can have a significant impact on the handwritten signature. Therefore, it is unrealistic to perform feature analysis on the signature images of each individual writer. This paper aims to propose a generic, writer-independent feature extraction method for offline handwritten signature. Firstly, the convolutional neural network model is pre-trained using a large public signature dataset to ensure that the model can learn sufficiently rich signature features and improve its generalization ability. In addition to the deep convolutional features extracted by the network, we also extracted different scale feature maps between multiple layers of the network to obtain feature information at different scales and levels. After fusing all feature maps of the same scale, we extracted high-dimensional local texture dense features to further capture the detailed texture information of the signature. Finally, factor analysis is used to reduce fusion feature dimensionality, reduce computational complexity, and preserve the most useful feature information, thereby improving the accuracy of recognition. Experiments are conducted on CEDAR and MCYT publicly available Latin datasets. The results demonstrate the effectiveness of the proposed feature extraction method, especially in cases with small training samples.</p>