Recently, convolutional neural network (CNN) compression via low-rank decomposition has achieved remarkable performance. Finding the optimal rank is a crucial problem because rank is the only hyperparameter for controlling computational complexity and accuracy in compressed CNNs. In this paper, we propose a global optimal rank selection method based on Bayesian optimization (BayesOpt), which is a machine learning based global optimization technique. By utilizing both a simple objective function and a proper optimization scheme, the proposed method produces a global optimal rank that provides a good trade-off between computational complexity and accuracy degradation. In addition, our method also reflects the correlation of each rank in multi-rank selection, and is able to flexibly yield an optimal rank with a given fixed compression ratio. Experimental results indicate that the proposed algorithm can identify the global optimal rank regardless of the huge size of dataset or the various structural features of CNNs. In all experiments on multi-rank selection, the proposed method produces the rank with higher accuracy and lower computational complexity than the state-of-the-art rank selection method, variational Bayesian matrix factorization (VBMF).

Anomaly detection is essential for many real-world applications, such as video surveillance, disease diagnosis, and visual inspection. With the development of neural networks, many neural networks have been used for anomaly detection by learning the distribution of normal data. However, they are vulnerable to distinguishing abnormalities when the normal and abnormal images are not significantly different. To mitigate this problem, we propose a novel loss function for one-class anomaly detection: decentralization loss. The main goal of the proposed method is to cause the latent feature of the encoder to disperse over the manifold space, such that the decoder can generate images similar to those in a normal class for any input. To this end, a decentralization term designed based on the dispersion measure for latent vectors is also added to the existing mean-squared error loss. To design a general solution for various datasets, we restrict the latent space by designing a decentralization loss term-based upper bound of the dispersion measure. As intended, a model trained with the proposed decentralization loss function disperses vectors on the manifold space and generates constant images. Consequently, the reconstruction error increases when the given test image is unknown. Experiments conducted on various datasets verify that the proposed function improves detection performance improved by about 1 % while reducing training time by 48 %, without any structural changes in the conventional autoencoder.

Single-image super-resolution technology has been widely studied in various applications to improve the quality and resolution of degraded images acquired from noise-sensitive low-resolution sensors. As most studies on single-image super-resolution focused on the development of deep learning networks operating on high-performance GPUs, this study proposed an efficient and lightweight super-resolution network that enables real-time performance on mobile devices. To replace the relatively slow element-wise addition layer on mobile devices, we introduced a skip connection layer by directly concatenating a low-resolution input image with an intermediate feature map. In addition, we introduced weighted clipping to reduce the quantization errors commonly encountered during float-to-int8 model conversion. Moreover, a reparameterization method was selectively applied without increasing the cost in terms of inference time and number of parameters. Based on the contributions, the proposed network has been recognized as the best solution in Mobile AI & AIM 2022 Real-Time Single-Image Super-Resolution Challenge with PSNR of 30.03 dB and NPU runtime of 19.20 ms.

The H.264/AVC video coding standard uses the rate distortion optimization (RDO) method to improve the compression performance in the intra prediction. Whereas the computational complexity is increased comparing with previous standards due to this method, even though this standard selects the best coding mode for the current macroblock. In this paper, a fast intra mode decision algorithm for H.264/AVC encoder based on dominant edge direction (DED) is proposed. The algorithm uses the approximation of discrete cosine transform (DCT) coefficient formula. By detecting the DED before intra prediction, 3 modes instead of 9 modes are chosen for RDO calculation to decide the best mode in the 4 times 4 luma block. For the 16 times 16 luma and the 8 times 8 chroma, instead of 4 modes, only 2 modes are chosen. Experimental results show that the computation time of the proposed algorithm is decreased to about 71% of the full search method in the reference code with negligible quality loss.

Robust principal component analysis, which extracts low-dimensional data from high-dimensional data, can also be regarded as a source separation problem of the sparse error matrix and the low-rank matrix. Until recently, various methods have attempted to precisely predict the discrete rank function by assigning a weight to the nuclear norm. However, if the weights are not in ascending order, the algorithms will diverge and exhibit high computational complexity. Moreover, from the viewpoint of source separation, these methods overlook the fact that two components must be sufficiently different for accurate demixing. In this paper, we employ the incoherence term with convex shape, which considers that components must appear different from one another for boosting separability. Since it is intractable to directly exploit mutual incoherence defined in linear algebra, we guarantee the incoherence by indirectly making the sparse matrix lack the low-rank property by using the duality norm principle. This approach can also be associated with the null space. To analyze the results of the proposed algorithm geometrically, we measure the geodesic distance between the tangent spaces of the manifolds of two separate components. As this distance increases, the degree of dissimilarity of the two components is adequately assured; thus, separation succeeds. Furthermore, this paper is the first to provide insights into the relationship between source separation conditions and the derivatives of the nuclear norm and $L_{1}$ norm. Experiments are conducted on still image separation and background subtraction to confirm the superiority of the proposed methods both qualitatively and quantitatively.

Retinex theory represents the human visual system by showing the relative reflectance of an object under various illumination conditions. A feature of this human visual system is color constancy, and the Retinex theory is designed in consideration of this feature. The Retinex algorithms have been popularly used to effectively decompose the illumination and reflectance of an object. The main aim of this paper is to study image enhancement using convolution sparse coding and sparse representations of the reflectance component in the Retinex model over a learned dictionary. To realize this, we use the convolutional sparse coding model to represent the reflectance component in detail. In addition, we propose that the reflectance component can be reconstructed using a trained general dictionary by using convolutional sparse coding from a large dataset. We use singular value decomposition in limited memory to construct a best reflectance dictionary. This allows the reflectance component to provide improved visual quality over conventional methods, as shown in the experimental results. Consequently, we can reduce the difference in perception between humans and machines through the proposed Retinex-based image enhancement.