Automated visual inspection of dents on metal parts immediately after press working presents two primary challenges. The first challenge is the varying inclination of the measured surface from part to part, which can be a significant obstacle. The second challenge is the uneven distribution of machining oil on the surface, which further complicates the inspection process. To address these issues, a method has been proposed to expand the measurement range of surface inclination and reduce the impact of machining oil adhesion. This method involves capturing a series of images with varying illumination angles. The challenges were addressed by normalizing the acquired series of images in the direction of the illumination angle and then integrating them. The effectiveness of this method was confirmed through experiments.

We propose a method for acquiring accurate posture and position of both hands for multiple users even from a distance using a system that consists of an ultrafast mirror-drive pan-tilt camera and a wide-view camera. By performing multithreaded high-speed gaze control for each target, the ultrafast pan-tilt camera can function as multiple virtual telephoto pan-tilt cameras. A wide-view camera is used to detect the positions of the user's hands, and each of the virtual telephoto pan-tilt cameras is utilized to obtain the images around the hands with high spatial resolution. By using these images, it is possible to achieve high-accuracy hands skeleton detection and obtain accurate hand postures. In addition, by synchronizing the whole body posture with the accurate hand postures and by correctly integrating the information obtained between different cameras, it is possible to obtain accurate hand positions. In the experiment, we acquired hand postures from a user standing 5 meters away from the camera system and classified them using an algorithm focused on the extension motion of each finger. As a result, our proposed method achieved an accuracy of 81.0% in classifying hand postures.

Due to the decrease in the number of key agricultural workers and the increase in the farmed area of cherries per worker, research and development is needed to save laboring time in cherry processing. Cherry fruit processing is heavily affected by multiple diseases and damage by pests or wildlife, increasing the time required to process cherries. To automatically differentiate healthy cherries from damaged cherries, a dataset consisting of 5,706 images of healthy cherries and cherries with natural damage was constructed. Ground truth images were created to show the segmentation of the damaged area. Furthermore, to verify the quality of the dataset in this study, the anomaly detection library Anomalib is used for training nine different state-of-the-art anomaly detection models using image sizes of 256 and 320, grouping healthy cherries into a Normal class and unhealthy cherries into an Anomalous class. Finally, we report the results from the training using common machine learning evaluation metrics for the image and pixel level predictions on both classes. Results show that high prediction performance can be achieved in the current state of the dataset, promoting its further enhancement, and setting the base for other image datasets for developing agriculture.

Even colorless and transparent samples can be optically observed without staining by detecting refractive index differences, as in differential interference contrast microscopy or phase-contrast microscopy. Surface Plasmon Polaritons (SPPs), excited on metal surfaces by light, are also well-known for detecting such refractive index differences. By using white light as the source to excite SPPs, the light with color corresponding to the refractive index of the sample medium excites the SPPs. By observing these colors, we can optically observe the samples. However, the SPP spectra are broadened, reducing the sensitivity. This paper proposes combining an optical resonator with SPP excitation to narrow the SPP excitation spectrum. Inside the resonator, light waves are reflected and travel back and forth multiple times, which enhances the waves whose phases are matched. Since the phase progression depends on the color, the resonator exhibits color selectivity. Considering SPP excitation, resonator parameters are designed through phase simulations to achieve a single-color peak in the spectrum. As a result, SPP excitation spectra with a reduced spectral width of the peak are obtained.

Current techniques for assessing the accuracy of 3D point cloud maps are often computationally demanding and lack the capability to pinpoint regions with diminished accuracy. In this research, we introduce an effective approach for evaluating the accuracy of 3D point cloud maps intended for autonomous driving systems. Our method focuses on calculating the entropy of the road surface along the vehicle's trajectory. By employing a moving average of Mean Map Entropy (MME), we can automatically identify areas where the map accuracy has degraded while also reducing the computational load. Through our evaluation, we demonstrate that the proposed method effectively detects point cloud blurring and surface duplication caused by SLAM/MMS errors.

The mainstream approach for both instance- and category-level pose estimation tasks is to detect the object as a preprocessing step and to estimate the pose from only local information focused on the object. The appearance of the object varies, even if the pose is the same, due to perspective effects. This is one of the reasons for the difficulty in learning the model. The effect of perspective is an unique feature of the object's position. In this paper, we propose a method to convert this into a position cue, which is a feature for pose estimation. Experimental results show that the use of position cue improves the performance of both instance- and category-level pose estimation tasks by 1-5%.

Geometric simulations are typically used to represent modern machining processes. However, most strategies rely on three-dimensional geometric models, which represent movements as a series of discrete snapshots. Physical laws and real-time interpolations are often used to estimate the state of motion at specific moments. To address this limitation, this study introduces a four-dimensional (4D) geometric model that simulates the tool-workpiece contact process within the spatiotemporal space. This article presents the fundamental concept of representing tools and workpieces in 4D space, followed by the proposal and implementation of two 4D cutter-workpiece engagement methods.

This study employs Large Vision-Language Model that integrates images and text to learn individuals' preferences and the rationales behind their judgments from a small dataset. Existing methods for estimating preferences and impressions have handled these by quantifying them; however, quantified preferences and impressions are difficult to interpret and lack explainability. Therefore, this research aims to enhance the explainability of preferences by generating the rationale behind preference judgments. We collected individual datasets by gathering personal preferences ("preferable" or "not preferable") and their rationales in specific domains and attempted to implement this by further training the Large Vision-Language Model. Experiments confirmed that, despite the dataset's limitation to only 200 images, it is feasible to estimate individual preferences and generate their rationales.

In this study, we propose a methodology for determining the quality of products based on a few images of non-defective and defective products, along with descriptions that serve as criteria for judgment. Existing Large Vision-Language Models (LVLMs) have demonstrated high performance across a variety of tasks, yet they lack specialized knowledge required for visual inspection. To address this issue, we enhance the LVLM's domain-specific expertise through additional training with a diverse collection of images of non-defective and defective products gathered from the web. Moreover, by utilizing In-Context Learning (ICL), our approach enables inference on inspection images based on a few exemplar images of non-defective and defective products, along with their judgment criteria descriptions, thereby eliminating the need for collecting extensive training samples and training models for each product type as traditionally required. By integrating LVLM with ICL, our method introduces a novel approach to general visual inspection, demonstrating its utility.

The paper conducts empirical experiments in the continual pre-training method using both artificially generated images and real images, and proposes an approach to model parameter initialization in continual pre-training. For the synthetic pre-training dataset, we use the VisualAtom-1k from formula-driven supervised learning, and for the real-image pre-training dataset, we assign the publicly available ImageNet-1k. In our experiments, we employ the Vision Transformer as a recognition model. Compared to pre-training with a single dataset, significant performance improvements were observed with our proposed setting. Furthermore, by implementing conditional model parameter initialization during continual pre-training, additional performance enhancements can be confirmed.