An Adaptive Fusion Feature Extraction Algorithm Based on CNN.

Keypoint feature extraction algorithms based on manual feature rely on the professional domain knowledge of the designer, and usually perform well in depicting the detail of an image. However, these manual feature-based feature extraction algorithms tend to fail in obtaining the abstract information of the image, which results in low robustness of the algorithms. In view of this, this paper proposes an image feature extraction algorithm that adaptively integrates the low layer feature of CNN and the high layer feature of CNN. The proposed algorithm makes use of fact that the low layer feature can detect the detail of an image and the high layer feature can detect the abstract information of the image. Meanwhile, the similarity between the area descriptors of two feature points is employed to adaptively determine the weight of the high layer feature and that of the low layer feature in the integrated feature, so as to realize the trade-off between distinguishability and invariance of features, and make the integrated feature more robust. Experiments on HPatches, Oxford, and RDNIM datasets show that the proposed algorithm is not only robust to illumination changes, but also shows superior performance when it comes to challenging scenes such as large viewing angle changes and day and night matching.