Abstract: A chunked LBP feature texture extraction method based on increasing weights of facial regions is proposed to address the problem of incomplete or partial loss of facial fatigue information resulting in inadequate characterisation of the driver's fatigue state, leading to low accuracy of fatigue detection. A basic image dataset for fatigue recognition is constructed based on driver images under different lighting conditions, and the acquisition and generalisation of sample images in the dataset is completed through self-built datasets, pre-processing and data enhancement work. An 8×8 block-weighted LBP algorithm is proposed and used to extract the driver's facial feature texture from the dataset images, which is used as the input of the convolutional neural network for model learning and training. The experimental results show that the proposed algorithm is fast in feature extraction, taking only 0.01s, and the fatigue detection model has good recognition accuracy and generalisation ability, with an accuracy rate of 93.52%. The proposed algorithm not only retains the ability to characterise image texture changes, but also effectively reduces feature redundancy, providing a feasible method for the recognition of driver fatigue states.