Learning region-wise deep feature representation for image analysis

Abstract

Effective feature representation plays an important role in image analysis tasks. In recent years, deep features, instead of hand-crafted features, have become the mainstream of the representation in image analysis tasks. However, the existing deep learning methods always extract feature representations from the whole image directly. Such strategies concentrate on extracting global features, and tend to fail in capturing local geometric invariance and introduce noise information from regions of not interest. In this paper, we propose a novel region-wise deep feature extraction framework for promoting the local geometric invariance and reducing noise information. In our algorithm, object proposal is adopted to generate a set of foreground object bounding boxes, from which the pre-trained convolutional neural network model is adopted to extract region-wise deep features. Then, an improved vector of locally aggregated descriptors strategy with weighted multi-neighbor assignment is proposed to encode the local region-wise feature representations. The final feature representation is not restricted to the classification task, and can also be further quantized to hash codes for large-scale image retrieval. Extensive experiments conducted on publicly available datasets demonstrate the promising performance of our work against the state-of-the-art methods in both image retrieval and classification tasks.