Local phase quantization plus: A principled method for embedding local phase quantization into Fisher vector for blurred image recognition

Advances in computer vision and image processing technology have led to great success in image recognition when the images are clear. However, in real-world applications, images are often blurred due to factors such as atmospheric turbulence, object-camera relative motion, and focus. This imposes a great challenge on practical image recognition tasks. To improve the performance of burred image recognition, one main approach is to extract blur-insensitive visual descriptors. A well-established blur-insensitive texture feature, local phase quantization (LPQ), can achieve promising results, with a trade-off between effectiveness and efficiency. However, for complicated visual recognition tasks, its performance is still not satisfactory. To leverage the discriminative power of LPQ, we propose local phase quantization plus (LPQ+), which embeds LPQ into Fisher vector (FV) to acquire mid-level blurred image representation under the bag-of-words (BoW) model. To better fit FV, instead of using real and imaginary parts, as in LPQ, LPQ+ directly quantizes the local phases of the short-term Fourier transform (STFT) directly. This results in lower-dimensionality features, but stronger local pattern characterization power. LPQ+ is densely sampled for blurred image representation; a sliding window screens the image with vertical and horizontal strides. LPQ+s are then acquired from all resulting local regions. To better maintain spatial structure characteristics, the sliding window is divided into finer cells. After being FV-encoded, local LPQ+s are aggregated through sum-pooling to generate the image signature. A wide range of experiments on 5 challenging datasets of different types (textures, faces, scenes, clouds, and flower) demonstrate that LPQ+ significantly outperforms LPQ and other well-established visual features in discriminative power.