Image Feature Fusion and Fisher Coding based Method for CBIR

摘要

The paper proposed a new method for content-based image retrieval (CBIR) based on image feature fusion and fisher encoding (FV). Firstly, low-level image content features such as hue-saturation-value (HSV) histogram, uniform local binary patterns (LBP), Dual-Tree complex wavelet transform (DTCWT) are extracted based on image blocks. In contrast, high-level features are extracted by using the AlexNet convolutional neural network (CNN). The singular value decomposition (SVD) was applied to the LBP and DTCWT. Secondly, low-level features are fused using normalization and weights. Lastly, after using the FV encoding, the fused fisher vectors are used to measure the similarity of image pairs. The experimental results on the benchmark Corel-1k show that the accuracy on the top 10, 12, and 20 images returned are 93.4%, 92.8%, and 91.4%, respectively.