X-ray image target recognition and model lightweight deployment

Aiming at the constraints of computing power and storage resources of embedded devices, this paper puts forward an improved model lightweight scheme, mainly introducing model distillation and model quantization methods. After L2 loss knowledge distillation experiment based on intermediate features, the accuracy of YOLOv5s model is improved by 1.6%. Through asymmetric quantization experiment of PTQ INT8, the speed is improved by nearly 2.6 times when the model accuracy is only 2.3%, which meets the demand of real-time reasoning. After the quantitative perceptual training, the speed of the model is increased by 1.8 times with a loss of only 1.4%. The improved lightweight model is deployed in embedded devices, and its performance is tested. Under the condition that the accuracy of the original model is 97% on RK3588S, the reasoning speed is increased by 2.6 times.