Robust Template Matching via Pruning Deep Feature

2018 2nd International Conference on Artificial Intelligence Applications and Technologies

计算机科学

Wang, F.^1 ; Xiong, J.P.^1 ; Yin, J.^1 ; Zhang, X.M.^1

Zhejiang Dahua Technology CO., LTD, Zhejiang Province, Hangzhou, China^1

关键词: Application environment;    Benchmark datasets;    Computational capacity;    Convolutional kernel;    Feature engineerings;    Industrial standards;    Template matching method;    Traditional computers;   

Others  :  https://iopscience.iop.org/article/10.1088/1757-899X/435/1/012056/pdf
DOI  :  10.1088/1757-899X/435/1/012056

学科分类：计算机科学（综合）

来源: IOP

【 摘 要 】

The application environment with slightly diversity between template image and target image has been the mainstream in template matching over the past decade or so. This paper, however, will discuss template matching method in such scenarios with erratic weather, deformations, scaling etc. For the feature engineering, by choosing the appropriate layers in a CNN and pruning inefficient convolutional kernels in the layers we want, we construct a feature space that can be used to represent more complex features compared to the traditional computer vision feature engineering, such as corners, colours, edges etc. Meanwhile, the feature space and computational capacity can be greatly reduced by the pruning operation on convolutional kernels. For similarity measure, a distance penalty term on the feature of image patches will be added in the final score function to make our method robust to deformation and scaling. Furthermore, the key coefficients of penalty term have been opened so that our method can be adjusted based on the actual scenarios. Numerous experiments on the benchmark dataset are conducted accompanied by comparisons with a few recent proposed methods, e.g., BBS, DDIS, etc. The results have demonstrated well the robustness and accuracy of our method relative to the other methods. Note that, our method can reach the industrial standard with GPU acceleration.