Earticle

본 논문에서는 가변 크기 블록 기반의 새로운 얼굴 특징 표현 방법을 제안한다. 기존 외형 기반의 얼굴 표정 인식 방법들 은 얼굴 특징을 표현하기 위해 얼굴 영상 전체를 균일한 블록으로 분할하는 uniform grid 방법을 사용하는데, 이는 다음 두 가지 문제를 가지고 있다. 얼굴 이외의 배경이 포함될 수 있어 표정을 구분하는 데 방해 요소로 작용하고, 각 블록에 포함된 얼굴의 특징은 입력영상 내 얼굴의 위치, 크기 및 방위에 따라 달라질 수 있다. 본 논문에서는 이러한 문제를 해결하기 위해 유의미한 표정변화가 가장 잘 나타내는 블록의 크기와 위치를 결정하는 가변 크기 블록 방법을 제안한다. 이를 위해 얼굴의 특정점을 추출하여 표정인식에 기여도가 높은 얼굴부위에 대하여 블록 설정을 위한 기준점을 결정하고 AdaBoost 방법을 이 용하여 각 얼굴부위에 대한 최적의 블록 크기를 결정하는 방법을 제시한다. 제안된 방법의 성능평가를 위해 LDTP를 이용하 여 표정특징벡터를 생성하고 SVM 기반의 표정 인식 시스템을 구성하였다. 실험 결과 제안된 방법이 기존의 uniform grid 기 반 방법보다 우수함을 확인하였다. 특히, 제안된 방법이 형태와 방위 등의 변화가 상대적으로 큰 MMI 데이터베이스에서 기존 의 방법보다 상대적으로 우수한 성능을 보여줌으로써 입력 환경의 변화에 보다 효과적으로 적응할 수 있음을 확인하였다.

Most existing facial expression recognition methods use a uniform grid method that divides the entire facial image into uniform blocks when describing facial features. The problem of this method may include non-face backgrounds, which interferes with discrimination of facial expressions, and the feature of a face included in each block may vary depending on the position, size, and orientation of the face in the input image. In this paper, we propose a variable-size block method which determines the size and position of a block that best represents meaningful facial expression change. As a part of the effort, we propose the way to determine the optimal number, position and size of each block based on the facial feature points. For the evaluation of the proposed method, we generate the facial feature vectors using LDTP and construct a facial expression recognition system based on SVM. Experimental results show that the proposed method is superior to conventional uniform grid based method. Especially, it shows that the proposed method can adapt to the change of the input environment more effectively by showing relatively better performance than exiting methods in the images with large shape and orientation changes.

Recently, Implicit Neural Representations (INRs) have been gaining attention as an approach where neural networks learn a continuous function that takes coordinates as input and outputs the color values at those locations. Using INRs allows for reconstructing images of any size without constraints on spatial resolution. As a result, it has emerged as a promising method for super resolution, enabling a single neural network to represent images at all resolutions. However, existing research on INR-based super-resolution still lags behind other deep learning methods in terms of performance. This is because a single neural network, which takes uniform coordinate values as input, faces challenges in representing information of varying complexity across different regions of an image. Therefore, we propose a method to improve super-resolution performance by decomposing an image into variable-sized blocks so that each block has uniform complexity, regardless of the regional variation in complexity. The INR neural network then learns the image information of each block with uniform complexity. By alleviating differences in regional complexity, the neural network is able to learn regional information more stably and accurately, enabling optimal performance even in areas with diverse levels of complexity.

The block matching algorithm for the motion estimation is relatively simple to implement, and thus widely applied in image sequence coding such as H.261, MPEG- I and MPEG-2. Most techniques of the block matching method use fixed-size blocks for the motion estimation. And their success relies on the assumption that the motion within each block is uniform. But if the block size is increased to reduce the number of motion vectors for high data compression, the estimated image brings about many errors. In this paper, the variable-sized blocks are used to solve this problem. And the top down method is used to select the block size.