Gaussian-filtered Local Difference Pattern with kernel representation for person-independent facial expression recognition robust to noise and resolution

A practical facial expression recognition system should be person-independent and robust against factors such as noise and resolution. This paper handles these issues and offers a new descriptor called Gaussian-Filtered Local Difference Pattern (GLDP). The inter-pixel and inter-frequency relationships of facial texture structures are encoded by GLDP. These multi-frequency patterns are obtained from low-pass and high-pass filtered images. First, the image is filtered with appropriate Gaussian filters. Next, the feature maps are obtained by computing and fusing the local difference vectors of each pixel of the low- and high-frequency filtered images. Then, the emotion-related blocks are detected and only the histogram of the created map of these blocks is used to get the feature vectors. Finally, these feature vectors are fed to a kernel representation algorithm to classify facial expressions. The proposed descriptor is evaluated against 20 other descriptors on four different datasets in terms of recognition rate, feature extraction time, base length, and number of bits using various person-independent experiments. The robustness of our method at low resolutions and noise variations is also examined. In addition, our method is compared with recent state-of-the art methods including deep learning methods under various testing strategies. The results demonstrate the superior performance of the proposed method. Code will be made available at: https://github.com/Najm-Morteza/GLDP-KR