Pedestrian Re-identification Based on Image Enhancement and Over-fitting Solution Strategies
Yong Shan Ding · 2018
Because of its application value and research significance, pedestrian re-identification (re-ID) technology has become more and more popular at present. This article has studied two common problems that affect the re-identification rate of pedestrians. Firstly, the pedestrian data collected in a specific scene has more blurred and obscured pedestrian images. The second is that directly using the data set to train deep neural networks is prone to over-fitting, resulting in poor performance of trained models. These problems make the final pedestrian recognition rate lower. This article is mainly based on logarithmic enhancement, histogram equalization and MSRCR algorithm (LHEMR) to enhance the blurred pedestrian image in the data set. This paper uses the random erasing technique and fine tuning for the second problem (Fine-tune). The combination of network model strategy to solve the problem of direct training deep neural network prone to over-fitting, the experimental results show that the proposed method is very effective and can improve the overall pedestrian recognition rate.