| Deep learning has made great progress in the application of computer vision,especially in the fields of person re-identification and object detection.Both person re-identification and object detection need to consider the model size in many practical scenarios.EfficientNet provides a series of convolutional neural network models that use a composite scaling technique for the purpose of model parameter optimization,including the network depth,width,and resolution.This thesis devotes to the pedestrian re-identification and object detection based on EfficientNet,and the main contribution can be summerized as follows:1)To address the problem of large model size of existing SOTA schemes in the field of person re-identification,an EfficientNet-based person re-identification network model named EPRI-Net is proposed,which inherits the idea of the feature pyramid technique and has the characteristics of small model size and excellent performance.Experiments show that the proposed EPRI-Net achieves 90.2% of mAP and 96.1% of rank-1 on the Market1501 dataset.2)For person detection tasks in practical scenarios,an EfficientNet-based small-scale person object detection model,named EPRI-Det,is proposed.EPRI-Det is designed on the basis of EPRI-Net,and its feature pyramid branches are used to detect the location of person object detection and class.Experiments show that the mAP of EPRI-Det on the crew dataset is 73.2%,which is slightly lower than that of Yolov5s(75.6%),but the model size is 4MB,which is substantially lower than that of Yolov5s(7.2MB).3)A grounded application of EPRI-Det for the crew safety cloak and hat wearing task is deployed with the Rexchip embedded platform,and an embedded target detection scheme based on EPRI-Det for crew safety cloak and hat wearing is proposed.The experiments show that the proposed scheme has an mAP of 56.2% after training and an inference time of about 470 ms for a single image after deployment to the embedded platform. |
PDF Full Text Request