Contrastive learning has proven effective in training sequential recommendation models by incorporating self-supervised signals from augmented views. Most existing methods generate multiple views from the same interaction sequence through stochastic data augmentation, aiming to align their representations in the embedding space. However, users typically have specific intents when purchasing items (e.g., buying clothes as gifts or cosmetics for beauty). Random data augmentation used in existing methods may introduce noise, disrupting the latent intent information implicit in the original interaction sequence. Moreover, using noisy augmented sequences in contrastive learning may mislead the model to focus on irrelevant features, distorting the embedding space and failing to capture users' true behavior patterns and intents. To address these issues, we propose Intent-aware Diffusion with contrastive learning for sequential Recommendation (InDiRec). The core idea is to generate item sequences aligned with users' purchasing intents, thus providing more reliable augmented views for contrastive learning. Specifically, InDiRec first performs intent clustering on sequence representations using K-means to build intent-guided signals. Next, it retrieves the intent representation of the target interaction sequence to guide a conditional diffusion model, generating positive views that share the same underlying intent. Finally, contrastive learning is applied to maximize representation consistency between these intent-aligned views and the original sequence. Extensive experiments on five public datasets demonstrate that InDiRec achieves superior performance compared to existing baselines, learning more robust representations even under noisy and sparse data conditions.

Large Language Models (LLMs) offer efficient and accessible support for audio engineers. However, text-only assistance lacks intuitive interaction and often fails to meet the practical demands of digital design. This study explores how an LLM can provide more effective support for audiobook producers, especially novices. We present a GPT-4o-powered analysis tool that extracts soundscape elements from text and align sound effects with semantic cues. By fusing LLM-guided interpretation with interactive audio control, the system introduces a new paradigm for supporting creative design through both semantic and acoustic dimensions. We invited three experts to evaluate the system and conducted a design experiment with 26 participants. Compared to the traditional method, our approach significantly improved design accuracy and efficiency. This work highlights the implicit tension and synergy between LLM-assisted creation and conventional design thinking, offering practical insights into the development of more adaptive and intelligent support tools for future audiobook production.

In recommender systems, a mainstream bias (MS bias) exists, which degrades the fairness of recommender systems and may decrease user satisfaction. In other words, the recommendation quality for users who prefer mainstream items is higher than that for users who prefer niche items. The MS bias degrades fairness in the recommender system and might decrease user satisfaction. To address the MS bias, we first need to quantify the degree of bias. However, the mainstreamness of users (the MS score) is difficult to measure properly, and we reveal that previous MS scores are greatly confounded by the number of user interactions that are irrelevant to mainstreamness. To address this issue, we propose a novel MS score that is unaffected by the number of interactions. Specifically, we introduce a new similarity metric that differs from the conventional Jaccard similarity measure to eliminate the effect of interaction numbers. We confirm the validity of our proposed MS score through a simulation study that discriminates mainstream users from niche users under varying numbers of user interactions. The proposed unconfounded MS score enables the proper evaluation of MS biases and the selection of a recommendation model with less MS bias. We evaluate the MS biases of two popular recommendation models using seven real-world datasets. We demonstrate that MS biases exist in all datasets and that a better model differs depending on the datasets. Furthermore, we confirm that the previous MS scores fail to select recommendation models with small biases. The results show the effectiveness of our new MS score in improving recommendations for fairness among users.

This study measured the depth of cracks on the ground surface by stereo matching aerial images of the cracks captured using a drone equipped with real time kinematics-global navigation satellite system (RTK-GNSS) and a high-resolution compact digital camera. Existing crack depth measurement methods are time-consuming, expensive, and cannot measure a wide area. In comparison, the proposed method can measure the depth of cracks in a short time and over a wide area by stereo matching only two aerial images. In addition, the drone's movement extends the baseline length, which increases the resolution in the height direction, resulting in precise depth measurements. To demonstrate the effectiveness of the proposed method, we took aerial images of cracks made of Styrofoam using shooting equipment made of aluminum frame and measured their depth. We realized stable millimeter-order crack depth measurements and achieved millimeter-order height resolution of 1.404 millimeters, and achieved 0.14 percent and 0.13 percent errors for 5 centimeters and 15 centimeters deep cracks, respectively.

Transfer learning performance in convolutional neural networks (CNNs) depends on the selection of retraining and fixation layers. The conventional method requires source data, which increases the cost proportionally to the data’s size. To solve this problem, this study proposed a source-data-less layer importance index that indicates the layers that should be retrained. Consequently, using the maximum activated feature maps generated by the model itself instead of the activated feature maps, the proposed method eliminates the need for any source data. Further, the experimental results on the five image classification datasets demonstrated a strong positive correlation of the proposed layer importance index compared with the conventional method. In addition, the transfer learning method using the layer importance index significantly improved the average classification accuracy compared with fine-tuning. CNNs have become indispensable owing to various applications, and thus, this study aimed at further enhancing its capabilities is relevant to the present and coming times.

The quality of three-dimensional (3D) reconstruction algorithm Structure from Motion (SfM) is affected by the input image’s resolution and CMOS’s noise level. We propose a denoisable Super Resolution (SR) method to improve resolutions while denoising for SfM’s input images taken by a CMOS device, improving its performance on noisy images. The conventional deep learning SR algorithm does not consider denoising during the learning process. This results in the disability of simultaneously reducing noise and improving resolution. In our methods Add Noise before Downsampling (An-Ds) and Downsampling before Adding Noise (Ds-An), instead of expanding the training data, we extract the noise from a real-world noise dataset and selectively add it to low resolution (LR) images of the SR training set. Thus the SR algorithm can simultaneously improve resolution and reduce noise. Moreover, selectively adding noise also remain SR’s performance on clean images. We trained two representative SR algorithms (SRCNN and EDSR) using traditional and our designed methods to process both clean and noisy images. Without changing the SR network’s structure, improvements of 0.17 dB in Peak Signal Noise Ratio (PSNR) by Ds-An and 0.14 dB by An-Ds (approximately 20% of improvement in three years) were observed in noisy images’ experiments by EDSR. Meantime, there is only a little loss (less than 0.01 dB) on the clean images. The SfM’s results show a better reconstructed 3D model using our methods. Compared to non-preprocessing and conventional preprocessing, key metrics such as Mean Reprojection Error (MRE) reduced 51.9% and 12.6%, and 2D key-points matching rate improved 41.7% and 217%, respectively.

The performance of transfer learning in convolutional neural networks depends on the selection of which layers are to be learned again and which are not. Generally, layers selection is performed manually; however, as the number of layers increases, the layers selection process becomes increasingly difficult. Thus, we propose a method to select an effective layers in transfer learning automatically using a genetic algorithm. In the proposed method, a genotype representing which layers' weights are updated or fixed in transfer learning is considered, and we achieve efficient layers selection in the way that a genotype with high validation accuracy is survived during genotype selection. Experiments are performed using the InceptionV3 network that pre-trained ImageNet as source images with transfer learning to CIFAR-100 as target images. Experimental results demonstrate that the test data accuracy in an ensemble of models whose layers are selected by the genetic algorithm is 15% and 12% greater than that of models trained by from-scratch and fine-tuning, respectively. In general transfer learning approach, layers on the output side are selected as adjustable layers; however, it is found that the distribution of the selected layers as an effective adjustable layers obtained by the genetic algorithm extends to the entire network. Transfer learning using a genetic algorithm may successfully capture the characteristics of a convolutional neural network's structure.

The Capsule Network (CapsNet) is a deep learning model proposed for image classification that is robust to pose of change of objects in images. A capsule is a vector representing the position, size and presence of an object. However, with CapsNet, the number of capsules increases, depending on the number of classification classes, and learning is computationally expensive. Thus, we propose a method for reducing computational costs by enabling a single capsule to represent multiple object classes. To learn the distance between classes, we incorporate the ArcFace distance learning method in the error function. In a preliminary experiment, the distribution of capsules was visualised by principal component analysis to demonstrate the validity of the proposed method. Using the MNIST and CIFAR-10 datasets, as well as an the affine transformed dataset, we compare the accuracy and learning time of the original CapsNet and proposed method. The results demonstrate that accuracy is improved by 2.74% on the CIFAR-10 dataset, and the learning time is reduced by more than 19% in both datasets.

Transfer learning can train a neural network for a target task's small dataset using a source task's pre-trained network; however, catastrophic forgetting, where the knowledge of the pre-trained network disappears during transfer learning, is problematic in this setting. PathNet was proposed to address this problem. However, PathNet can only be applied to modular neural network cases:thus, a non-modular pre-trained neural network is unavailable. Consequently, PathNet cannot be used to improve network structure versatility. Therefore, we propose Stepwise PathNet to improve versatility by considering the layers of a non-modular pre-trained neural network as modules. The performances of the proposed Stepwise and original PathNet methods were compared using the CIFAR-10 dataset (10 classes, and 60,000 images), and the results confirm the proposed method's potential to stabilize learning curves and accelerate learning to 45%.

Despite the benefits of using drones in addressing the demand in agricultural production, the number of drone accidents is increasing rapidly, and accidents such as drone crashes and contacts may cause serious damage, Hence there is a need to address accidents associated with drone and how to avoid them, This paper focuses on propellers that reside inside the drone. Here, the propulsion mechanism of the proposed safe drone is based on the Coanda effect, and the purpose of this paper is to optimize this propusion mechansim. Regarding optimizing the design, there are multiple design variables and objective variables, and when this combination becomes enormous, an expensive fluid-simulation calculation is required to evaluate the performance of each design candidate. This paper proposes an effective design method that uses a self-organizing map to generate relatively few design candidates from which to select the optimum one. In evaluation experiments, 108 individual drones are created using three-dimensional computer-aided design, and a fluid simulator is used to calculate the propulsive force required for flight based on six design variables for the propulsion mechanism. We set the propulsive power and the weight of these candidates as the two objective variables, and we use the self-organizing map to visualize. The results show that the evaluation index of the optimized designs drones is roughly 7% better than that of the standard design.

We propose a superresolution (SR) algorithm based on adaptive sparse representation via multiple dictionaries for images taken by unmanned aerial vehicles (UAVs). The SR attainable through the proposed algorithm can increase the precision of 3-D reconstruction from UAV images, enabling the production of high-resolution images for constructing high-frequency time series and for high-precision digital mapping in agriculture. The basic idea of the proposed method is to use a field server or ground-based camera to take training images and then construct multiple pairs of dictionaries based on selective sparse representations to reduce instability during the sparse coding process. The dictionaries are classified on the basis of the edge orientation into five clusters: 0, 45, 90, 135, and nondirection. The proposed method is expected to reduce blurring, blocking, and ringing artifacts especially in edge areas. We evaluated the proposed and previous methods using peak signal-to-noise ratio, structural similarity, feature similarity, and computation time. Our experimental results indicate that the proposed method clearly outperforms other state-of-the-art algorithms based on qualitative and quantitative analysis. In the end, we demonstrate the effectiveness of our proposed method to increase the precision of 3-D reconstruction from UAV images.

The present paper generalizes the problems of nonnegative matrix factorization and decomposition of fuzzy relation into a common non-linear non-negative matrix factorization problem. Algorithms for solving such a general nonlinear problem are discussed, based on general algebraic structures of ordered semirings with generated pseudo-operations. Some decompositions in max-product, max-plus algebras are also shown.

Max-plus algebra based wavelet transforms are considered currently to be the world's fastest image processing methods. In the present paper a novel max-plus algebra based wavelet transform is proposed and studied both from the theoretical and practical points of view. The novel MP-wavelets considered here employ strictly neighbor pixels in the calculations and their most important properties are: very low computational complexity, flexible sampling window size and potentially, very easy hardware implementation.

Recently it has been shown that in image processing, the usual sum and product of the reals are not the only operations that can be used. Several other operations provided by fuzzy logic perform well in this application. We continue this line of research and we propose the use of a pair consisting of a uninorm and an absorbing norm determined by a continuous, strictly increasing generator instead of the classical sum and multiplication. In the present paper the discrete cosine transform (DCT)-based image compression method is generalized by using a pair consisting of a uninorm and an absorbing norm. We show that the results of the proposed method outperform in several cases the classical DCT image compression algorithm.

Recently it has been shown that in image processing, the usual sum and product of the reals are not the only operations that can be used. Several other operations provided by fuzzy logic perform well in this application. We continue this line of research and we study the possibility to use some pairs of pseudo-operations. We define in the present paper pseudo-linear Haar wavelets, and we perform multi-channel decomposition of images. We study some pairs of pseudo-operations determined by a continuous, strictly increasing generator instead of the classical sum and multiplication and Haar-type wavelets based on these operations. The results show us that pseudo-linear Haar wavelets can be used as an alternative of classical Haar wavelets since the perfect reconstruction property is conserved.

In order to visualize the whole structure of concept lattice obtained by formal concept analysis with respect to the image/video databases, a hierarchical representation method based on fuzzy clustering (especially, FCM) is proposed. In the case of proposed method, the FCM is firstly performed with respect to vast amount of objects of image/video databases, and the number of objects is reduced into suitable numbers for visualization under the environmental constrains. Each object of updated information table corresponds to each centroid of cluster. Furthermore, a flexible formal concept analysis can be formulated since the FCM can treat with multiple cluster information with respect to each object. Experiments using real image/video databases ('Corel Gallery 1000a' (1,000 color images) and a video selected from standard motion database (150 frames)) are performed to confirm the effectiveness of the proposed method. Through the investigation of the concept lattice obtained by the proposed method, it is confirmed that the proposed method is helpful to do video clipping and grasp the whole structure of image database.

Recently it has been shown that sum and product are not the only operations that can be used in order to define concrete approximation operators but several operations provided by fuzzy logic can be used. In this sense, in the present paper, pseudo-linear approximation operators of Shepard-type are studied from the practical point of view, in function approximation. It is shown that in several cases these outperform classical approximation operators based on sum and product operations.

A wavelet transformation based on max-plus algebra is proposed, where analysis and synthesis operations are defined by max (or min) and standard sum and they treat with 9 channel multi resolution images. The proposed wavelets especially can reserve edge information on compressed images due to the non-linear operations (max and min); therefore, they are efficient for the compression of predictive frames used in video compression and reconstruction. Through video compression and reconstruction experiments using standard video database, it is confirmed that effectiveness of the proposed wavelets.

Generalized non-linear wavelets are proposed by using weight coefficients in the analysis/synthesis operations. The proposed wavelets can generate various types of mother wavelets by adjusting the weight coefficients, in the setting of non-linear signal processing. Through experiment using real images extracted from standard image database (SIBDA), image analysis results (sub-band decomposition) are represented by the proposed method. Furthermore, the proposed method is applied to preprocessing of the chest lung X-ray images diagnosis based on standard digital image database created by Japanese society of radiological technology, and results of suspicious regions detection are shown

An image can be interpreted as a fuzzy relation by normalizing the values of its pixels. We use the greatest eigen fuzzy set (for short, GEFS) with respect to the max-min composition and the smallest eigen fuzzy set (for short, SEFS) with respect to the min-max composition of this fuzzy relation for resolution of problems of image information retrieval. The experiments are executed on some images extracted from "view sphere database". Based over GEFS and SEFS, a similarity measure is introduced for comparison between the sample image and the retrieved images.

In order to generate various eigen fuzzy sets with respect to an image, an algorithm based on a permutation fuzzy matrix is proposed. The proposed algorithm is based on a rotation of the original fuzzy relation by using permutation fuzzy matrix, and the periodicity of the permutation matrix is studied in the setting of the fuzzy relational calculus. Through an experiment using the images extracted from the 'view sphere database', the properties of various eigen fuzzy sets obtained by the proposed algorithm are shown.

A color motion compression/reconstruction method based on fuzzy relational equations (C-MCF) is proposed. The C-MCF divides a motion sequence into intra-pictures (I-pictures) and predictive-pictures (P-pictures), respectively, and these pictures furthermore are divided into R, G, and B pictures that are equivalent to fuzzy relations. The fuzzy relations are compressed/reconstructed by fuzzy relational equations, where I-pictures and P-pictures are compressed/reconstructed by using uniform and non-uniform coders, respectively. In order to perform efficient compression/reconstruction of P-pictures, a design method of non-uniform coders is proposed. The proposed design method is based on a fuzzy entropy function and an overlap level of fuzzy sets and a fuzzy equalization. An experiment using 10 P-pictures confirms that the root means square errors of the proposed method is decreased to 77.2% of that of the uniform coders, under the compression rate 0.00198. An experiment of motion compression and reconstruction is also presented to confirm the effectiveness of the C-MCF based on the non-uniform coders.

A digital watermarking method using image compression based on a fuzzy relational equation (ICF) is proposed. The method is based on least significant bit modification. If the coding system of ICF is not designed appropriately, the fuzzy relational equation will be unsolvable due to the watermarking (modification of compressed image). In order to avoid this problem, a condition for appropriate coding system design is represented in terms of the solvability degree of the fuzzy relational equation. Image compression and reconstruction experiments using 100 images (extracted from Corel Gallery) are performed, and it is confirmed that the signed image is indistinguishable from the unsigned one.

A fast solving method of the solution for max continuous t-norm composite fuzzy relational equation of the type G(i, j)=(R/sup T//spl square/A/sub i/)/sup T//spl square/B/sub j/, i=1, 2, ..., I, j=1, 2, ..., J, where A/sub i//spl isin/F(X)X={x/sub 1/, x/sub 2/, ..., x/sub M/}, Bj/spl isin/F(Y) Y={y/sub 1/, y/sub 2/, ..., y/sub N/}, R/spl isin/F(X/spl times/Y), and /spl square/: max continuous t-norm composition, is proposed. It decreases the computation time IJMN(L+T+P) to JM(I+N)(L+P), where L, T, and P denote the computation time of min, t-norm, and relative pseudocomplement operations, respectively, by simplifying the conventional reconstruction equation based on the properties of t-norm and relative pseudocomplement. The method is applied to a lossy image compression and reconstruction problem, where it is confirmed that the computation time of the reconstructed image is decreased to 1/335.6 the compression rate being 0.0351, and it achieves almost equivalent performance for the conventional lossy image compression methods based on discrete cosine transform and vector quantization.