Research On Semantic Segmentation Based On Quantization And Knowledge Distillation

With the development and popularization of the fifth-generation mobile communication technology,the vision of the intelligent interconnection of all things is getting closer and closer to the daily life of human beings.Thanks to the technical characteristics of 5G technology’s ultra-high speed,ultra-low latency,and large connection volume,embedded devices have a broader space for development.Generally speaking,the application of deep learning technology is an important way to realize the intelligence of embedded devices.However,it is very difficult for embedded devices to run deep learning models with hundreds or even tens of millions of parameters.Due to cost,size and energy consumption,the computing power of embedded devices is often difficult to match the parallelism of multiple cards.Therefore,how to improve the prediction accuracy of the lightweight model is very important.At present,the demand for semantic segmentation tasks in embedded scenes is very common,such as intelligent driving,remote sensing images,and scene understanding.The existing research on lightweight semantic segmentation often starts from the structure.Although this method can reduce the number of parameters of the model,the design process is complicated and the application method lacks flexibility.In view of the above problems,the main work of this thesis is as follows:First.Based on model quantization and knowledge distillation technology,a semantic segmentation method based on quantization and knowledge distillation is proposed,which can well improve the performance of lightweight models,and has very good flexibility,which can be easily embedded into existing networks.in the training process.Using the semantic segmentation method of quantization and knowledge distillation,a lightweight student network based on residual structure is trained on the Cityscapes dataset,and the mIoU index of 77.56%is achieved on the validation set,The feasibility of the semantic segmentation method based on quantization and knowledge distillation is verified.Second,Based on the further optimization of the above student network,a more lightweight semantic segmentation network is constructed and named Mob student.Depth separable convolution is used to replace convolution operation to reduce the amount of parameters,Inverted Residual Block is used to replace Bottleneck structure to improve network feature extraction capability,and the same training strategy is used to train the improved student network,which further improves network operation and storage efficiency.Third,The performance of each model is tested based on the Nvidia Jetson TX1 embedded device.The control variable method is used to test the forward inference speed of the residual structure-based student network and Mob student on embedded devices.The same model was independently tested for 100 times,and finally the average time-consuming of the student network based on residual structure and the forward inference based on Mob student on Nvidia Jetson TX1 were 0.74 seconds and 0.34 seconds,respectively,Test the actual effect of the above network on embedded devices.