Automatic classification of digital patent images is significant for improving the efficiency of patent examination and management. In this paper, we propose a new patent image classification method based on an enhanced deep feature representation. Convolutional neural networks (CNN) is novelly applied to the patent image classification. The synergy between deep learning and traditional handcraft feature is explored. Specifically, the deep feature is first learned from massive patent image samples by AlexNet. Then such deep learning feature is further enhanced by fusing with two kinds of typical handcraft features including local binary pattern (LBP) and adaptive hierarchical density histogram (AHDH). In order to obtain a more compact feature representation, dimension of the fused feature is subsequently reduced by PCA. Finally, the patent image classification is conducted by a series of SVM classifier. Statistical test results on a large-scale image set show that the state-of-the-art performance is achieved by our proposed patent image classification method.
Proc. SPIE. 10033, Eighth International Conference on Digital Image Processing (ICDIP 2016)
KEYWORDS: Principal component analysis, Image compression, Visual process modeling, Digital image processing, Visualization, Computer programming, Associative arrays, Image classification, Performance modeling, Current controlled current source
Large visual dictionaries are often used to achieve good image classification performance in bag-of-features (BoF) model, while they lead to high computational cost on dictionary learning and feature coding. In contrast, using small dictionaries can largely reduce the computational cost but result in poor classification performance. Some works have pointed out that pooling locally across feature space can boost the classification performance especially for small dictionaries. Following this idea, various pooling strategies have been proposed in recent years, but they are not good enough for small dictionaries. In this paper, we present a unified framework of pooling operation, and propose two novel pooling strategies to improve the performance of small dictionaries with low extra computational cost. Experimental results on two challenging image classification benchmarks show that our pooling strategies outperform others in most cases.