The performance of Hand Gesture Recognition(HGR)depends on the hand shape.Segmentation helps in the recognition of hand gestures for more accuracy and improves the overall performance compared to other existing deep n...The performance of Hand Gesture Recognition(HGR)depends on the hand shape.Segmentation helps in the recognition of hand gestures for more accuracy and improves the overall performance compared to other existing deep neural networks.The crucial segmentation task is extremely complicated because of the background complexity,variation in illumination etc.The proposed mod-ified UNET and ensemble model of Convolutional Neural Networks(CNN)undergoes a two stage process and results in proper hand gesture recognition.Thefirst stage is segmenting the regions of the hand and the second stage is ges-ture identification.The modified UNET segmentation model is trained using resized images to generate a cost effective semantic segmentation model.The Central Processing Unit(CPU)utilization and training time taken by these models with respect to three public benchmark datasets are also analyzed.Recognition is carried out with the ensemble learning model consisting of EfficientNet B0,Effi-cientNet B4 and ResNet V2152.Experimentation on NUS hand posture dataset-II,OUHANDS and HGRI benchmark datasets show that our architecture achieves a maximum recognition rate of 99.07%through semantic segmentation and the Ensemble learning model.展开更多
文摘The performance of Hand Gesture Recognition(HGR)depends on the hand shape.Segmentation helps in the recognition of hand gestures for more accuracy and improves the overall performance compared to other existing deep neural networks.The crucial segmentation task is extremely complicated because of the background complexity,variation in illumination etc.The proposed mod-ified UNET and ensemble model of Convolutional Neural Networks(CNN)undergoes a two stage process and results in proper hand gesture recognition.Thefirst stage is segmenting the regions of the hand and the second stage is ges-ture identification.The modified UNET segmentation model is trained using resized images to generate a cost effective semantic segmentation model.The Central Processing Unit(CPU)utilization and training time taken by these models with respect to three public benchmark datasets are also analyzed.Recognition is carried out with the ensemble learning model consisting of EfficientNet B0,Effi-cientNet B4 and ResNet V2152.Experimentation on NUS hand posture dataset-II,OUHANDS and HGRI benchmark datasets show that our architecture achieves a maximum recognition rate of 99.07%through semantic segmentation and the Ensemble learning model.
