目前,无监督单模态行人重识别研究主要集中于可见光图像。随着新型红外摄像头的普及,无监督红外行人重识别也展现出其研究价值。由于红外图像对比度低、缺乏颜色纹理细节信息,因此全局信息对于红外行人重识别至关重要。本文设计了基于F-...目前,无监督单模态行人重识别研究主要集中于可见光图像。随着新型红外摄像头的普及,无监督红外行人重识别也展现出其研究价值。由于红外图像对比度低、缺乏颜色纹理细节信息,因此全局信息对于红外行人重识别至关重要。本文设计了基于F-ResGAM的无监督红外行人重识别网络。该网络首先利用小波变换对图像进行预处理以增强特征提取能力,接着在resnet50网络结构中引入全局注意力机制(Global Attention Mechanism,GAM)关注更多的全局信息。此外,由于红外伪标签噪声较大,本文提出采用基于样本扩展的分组采样(Group Sampling based on Sample Expansion,GSSE)策略进一步优化伪标签生成,从而提升了模型的识别精度。实验结果表明,本文提出的优化方法有效提升了无监督红外行人重识别的精度,尤其是rank指标显著提升。展开更多
We studied the variation of image entropy before and after wavelet decomposition, the optimal number of wavelet decomposition layers, and the effect of wavelet bases and image frequency components on entropy. Numerous...We studied the variation of image entropy before and after wavelet decomposition, the optimal number of wavelet decomposition layers, and the effect of wavelet bases and image frequency components on entropy. Numerous experiments were done on typical images to calculate (using Matlab) the entropy before and after wavelet transform. It was verified that, to obtain minimal entropy, a three-layer decomposition should be adopted rather than higher orders. The result achieved by using biorthogonal wavelet decomposition is better than that of the orthogonal wavelet decomposition. The results are not directly proportional to the vanishing moment, however.展开更多
For any fixed ε > 0, an explicit construction of anorthonormal trigonometric polynomial basis {Tk}∞k=1 inL2 [0,1) with degTk≤ 0.5(1 +ε)k is presented. Thus weimprove the results obtained by D. Offin and K. Osko...For any fixed ε > 0, an explicit construction of anorthonormal trigonometric polynomial basis {Tk}∞k=1 inL2 [0,1) with degTk≤ 0.5(1 +ε)k is presented. Thus weimprove the results obtained by D. Offin and K. Oskolkov in [4] and by Al. A. Privalov in [6]. and practically solve the open problemasked in [4], [8] and [9]. Moreover, as in [4], Fourier sums with respectto this polynomial basis are projectors onto subspaces of trigonometricpolynomials of high degree, which implies almost best approximation- properties.展开更多
文摘目前,无监督单模态行人重识别研究主要集中于可见光图像。随着新型红外摄像头的普及,无监督红外行人重识别也展现出其研究价值。由于红外图像对比度低、缺乏颜色纹理细节信息,因此全局信息对于红外行人重识别至关重要。本文设计了基于F-ResGAM的无监督红外行人重识别网络。该网络首先利用小波变换对图像进行预处理以增强特征提取能力,接着在resnet50网络结构中引入全局注意力机制(Global Attention Mechanism,GAM)关注更多的全局信息。此外,由于红外伪标签噪声较大,本文提出采用基于样本扩展的分组采样(Group Sampling based on Sample Expansion,GSSE)策略进一步优化伪标签生成,从而提升了模型的识别精度。实验结果表明,本文提出的优化方法有效提升了无监督红外行人重识别的精度,尤其是rank指标显著提升。
基金the Natural Science Foundation of China (No. 60472037).
文摘We studied the variation of image entropy before and after wavelet decomposition, the optimal number of wavelet decomposition layers, and the effect of wavelet bases and image frequency components on entropy. Numerous experiments were done on typical images to calculate (using Matlab) the entropy before and after wavelet transform. It was verified that, to obtain minimal entropy, a three-layer decomposition should be adopted rather than higher orders. The result achieved by using biorthogonal wavelet decomposition is better than that of the orthogonal wavelet decomposition. The results are not directly proportional to the vanishing moment, however.
文摘For any fixed ε > 0, an explicit construction of anorthonormal trigonometric polynomial basis {Tk}∞k=1 inL2 [0,1) with degTk≤ 0.5(1 +ε)k is presented. Thus weimprove the results obtained by D. Offin and K. Oskolkov in [4] and by Al. A. Privalov in [6]. and practically solve the open problemasked in [4], [8] and [9]. Moreover, as in [4], Fourier sums with respectto this polynomial basis are projectors onto subspaces of trigonometricpolynomials of high degree, which implies almost best approximation- properties.
