Ssd-kdgan: a lightweight SSD target detection method based on knowledge distillation and generative adversarial networks

Usually, high-accuracy target detection algorithms have many model parameters, which require much storage space and a long testing time and are difficult to deploy on intelligent inspection robots with limited storage capacity and computational resource capability. Although lightweight target detection algorithms have lower model parameters, they cannot meet the demand for high-precision detection. To solve this problem, we design a lightweight target detection method SSD-KDGAN using knowledge distillation and generative adversarial networks. SSD-KDGAN transfers the feature maps generated by deeper and more complex teacher networks as shared knowledge to a student model with a small number of parameters without increasing testing costs, to improve the detection accuracy of the lightweight student model. First, we align the feature maps generated by the student model and the teacher model to avoid information loss caused by forced training between feature maps of different scales. Second, we use the knowledge of the teacher network to guide the learning of the student network through distillation loss. Finally, we make the feature maps output from the student network have similar spatial distributions as the feature maps output from the teacher network through adversarial loss. Extensive experiments on the PASCAL VOC and images in real environments validate the effectiveness of our framework. The experimental results indicate that transferring the knowledge of the teacher network to the student network can significantly improve the detection accuracy of the student network without increasing the testing cost.