To address the issue that the existing multitarget detection network cannot extract and enhance dynamic flame features, resulting in poor detection results, this paper presents an improved YOLOv3 flame detection algorithm based on dynamic shape feature extraction and enhancement. ResNet50_vd with a small size structure is used as the backbone network of YOLOv3 to reduce the redundancy of feature information. To control the dynamic change of the sampling grid with the shape of the flame target, deformable convolutional neural network modules are added to the backbone network stage 4 and stage 5. The IoU Aware module is introduced to increase the correlation between the confidence score and the positioning accuracy of the IoU, and to enhance the flame feature extraction ability of the network. Simultaneously, the Drop Block module is added to the YOLOv3 Head, and the IoU prediction component is introduced to optimize the loss function, which improves the feature enhancement ability during the model learning process. The ablation experiments were performed to verify the effect of each improvement on the proposed model. The results show that the improved model for flame detection has a detection accuracy of 94.11% and an inference speed of 73.52 frame/s, which can effectively meet the detection requirements of dynamic shape flames.