Abstract: We solved the problem of low accuracy in car window recognition of the original YOLOX-S model by introducing deformable convolutional neural networks and Focal loss function (Focal loss) to the YOLOX-S model. Firstly, by introducing deformable convolutional neural networks into the backbone feature extraction network of the YOLOX-S model, offsets were introduced for each sampling point in the convolutional kernel to facilitate the extraction of more representative information from the original image, thereby improving the accuracy of car window recognition. Secondly, using Focal loss instead of binary cross entropy loss function in the original model, Focal loss could alleviate the impact of imbalance between positive and negative samples on training, and it paid  more attention to difficult samples during the training process, thereby improving the  recognition performance of the model for car window targets. Finally, in order to verify the performance of the improved algorithm, 15 627 images were collected and annotated for training and validation in the experiment. The experimental results show that the average target accuracy of the improved car window recognition algorithm increases by 3.88%.

Key words:  car window recognition, YOLOX-S model, deformable convolutional neural network, Focal loss