SHS-ReID: A Person Re-Identification Method Based on Semi-Hard Sample Mining and DDT Neural Network Framework

Person re-identification (ReID) task is complex because it usually faces many challenges, such as view changes, pose changes, and occlusions. To improve the accuracy of ReID, we propose a method based on semi-hard sample mining and a novel DDT (deferred, decompose and transpose) neural network framework. First, we obtain the DDT neural network framework through deferring downsampling, decompose large convolution kernel and step size transpose operations, of Resnet-baseline network model, and then transferred person images into the DDT framework to further obtain global features and local features. The DDT framework obtains more person feature information, which can enhance the accuracy of ReID. Finally, we add semi-hard sample mining strategy before triplet loss classification to optimize the ability of selecting samples. Experimental results show that the proposed DDT neural network framework can reduce the calculation amount of the relevant layer by 45%, meanwhile the ability of up-taking feature is increased by 75% in the relevant layer. Additionally, the accuracy of ReID is improved by 1% – 2% in overall. The experimental results verify the effectiveness of the semi-hard sample mining strategy and the DDT framework.