Design And Implementation Of Vespa Real-time Recognition System On Mobile Phone Based On Deep Learning

Deep learning has been applied to many fields for its rapid development,but the field of vespa recognition still lacks specific automatic recognition algorithms and lacks a comprehensive and systematic data set.In this thesis,we propose a real-time fine-grained recognition algorithm for Vespas,which mainly includes:basic feature extraction module,insert non-local into the ResNet-vd network,and improve the basic network’s ability to extract high-level semantic features without losing speed as much as possible;The force module introduces an attention mechanism in the two dimensions of channel and space to improve the feature representation and feature extraction capability of the model;the feature map sampling module uses the attention map to sample the original image,highlights the discrimination area with high resolution,and focuses on different discrimination areas.The attention map loop iteratively refines the image details asynchronously;the feature fusion module uses bilinear pooling to fuse the original image feature map and the attention map to obtain the feature matrix.We also design a classification loss function that includes the center loss of the attention map of the same channel in the same category to focus on a specific field to improve the classification accuracy.In order to make the application of the model faster on the mobile terminal,we combine cutting and knowledge distillation to compress the model and increase the speed of the model.The comparative experiments between the model and other fine classification algorithms prove that the method proposed in this thesis is superior to the other algorithms for vespa recognition.Finally,we build a mobile terminal system equipped with a vespa recognition algorithm,which provides users with basic functions such as identifying vespa and uploading data.The system is reliable enough for practical application.