Deep learning semantic segmentation model compression method for embedded mobile terminal

Abstract

The invention discloses a deep learning semantic segmentation model compression method for an embedded mobile terminal. According to the invention, teacher network parameter weight obtained by training is fixed. Identification network and student network are continuously trained and learned. Distillation (paired distillation, pixel distillation and integral distillation) is performed on three different levels. An overall optimization target (cross entropy loss, pixel distillation loss, paired distillation loss and overall distillation loss) is continuously optimized. Finally, the number of parameters of the student network obtained through distillation is greatly reduced and the forward calculation time of the network is greatly reduced under the condition that the student network obtainedthrough distillation meets extremely small reduction of IEU (Interoperability Union). The problem that the embedded mobile terminal cannot carry a large deep learning network under the condition thatGPU capacity and power supply are limited is solved. The task calculation time is greatly shortened and the embedded mobile terminal platform carries a complex deep network model.

Description

technical field [0001] The invention relates to the field of automatic driving, especially the deep learning semantic segmentation task on the embedded mobile terminal platform, and provides a compression method for the semantic segmentation model of the embedded mobile terminal, so that the embedded mobile terminal platform can calculate parameters when performing the semantic segmentation task Less amount, faster calculation. Background technique [0002] In recent years, convolutional neural network, as a commonly used basic network module in deep learning, has gradually become indispensable in tasks such as image classification, target detection and semantic segmentation in computer vision tasks. [0003] Since the recognition accuracy of deep learning algorithms in the field of target recognition was improved in 2012, various deep learning algorithms have shined in different computer vision tasks. As these deep learning algorithm models continue to approach the limit o...

AI Technical Summary

Problems solved by technology

[0004] Due to its weak heat dissipation capability and cost sensitivity, the mobile platform cannot be equipped with a higher-level graphics processor, and the parameters of the graphics processor determine the upper limit of the parameters of the network model that it can run; at the same time, the depth of the model also affects the model. Computing time, a deeper network will generate more intermediate variables during the training phase, these intermediate variables consume computing time during the backpropagation process, and also occupy a fixed memory space

Images