Research On Semantic Segmentation Technology Based On Deep Learning In Android Mobile Platform

Semantic segmentation is a basic task in the field of computer vision,which assigns a label to each pixel of an image to provide pixel-level fine recognition results.Although current mainstream semantic segmentation algorithms have high precision,they contain a lot of parameters and calculations,which limits the model being deployed on the edge computing platform.Therefore,this paper proposes a semantic segmentation algorithm suitable for Android mobile platform.This paper focuses on semantic segmentation algorithm from three aspects:model training strategy optimizing,model structure improving and model lightening.The robustness and generalization of the model are improved by model pre-training and auxiliary supervision.A new semantic segmentation network named SemSegFPA is proposed,which uses attention modules to enhance the ability of understanding global context and recognizing edges.The model is lightened by replacing the backbone with MobileNetV2 network,and the inferece speed of the model is nearly doubled.We deployed the improved semantic segmentation algorithm SemSegFPA to Android platform and tested its performace.Experiments shows that SemSegFPA algorithm can smoothly run on Mi 8 mobilephone,that possess CPU Qualcomm Snapdragon 845 hardware configuration and 6GB running memory,with inference time approximately 1～2 second.Compared with current mainstream semantic segmentation algorithms,the improved algorithm we proposed in this paper has less computation,and it can smoothly run on Android mobile phones with a good performance and faster inference speed,which shows the practicability of our study.