The work aims to solve the problems such as the difficulty of stacking classification and the low handling efficiency caused by the non-uniform size and specification of customized wooden doors and the diversity of surface textures. A deep learning method was proposed to detect customized wooden door workpieces, and a robot workpiece recognition method was studied based on YOLO V3 network. First, through image data enhancement and preprocessing, the customized wooden door data were expanded. Then, the YOLO V3 loss function was improved, and the anchor frame scale of the customized wooden door data set was re-clustered according to the characteristics of the wooden doors. Finally, the spatial pyramid pooling layer was applied to improve the feature pyramid network in YOLO V3, and the effectiveness of this method was verified by a randomly selected test set. The average detection accuracy of the test data set reached 98.05%, and the detection time of each image was 137 ms. The research shows that this method can meet the requirements of the wooden door production line for accuracy and real-time nature, and can greatly improve the turning line and stacking efficiency of customized wooden doors.