In recent years, attempts to detect video anomalies in daily life, manufacturing sites, etc. using deep learning have been investigated. In particular, generative adversarial networks (GANs) have been widely used. However, they are inefficient, and simple difference-based methods are affected by noise. This paper proposes a new unsupervised learning method for video anomaly detection using spatio-temporal generative adversarial networks. Since the proposed method utilizes the middle layer features of the U-Net Discriminator, it is able to detect anomalies efficiently and accurately in the region of interest of the Discriminator.

We investigated how the observer's gaze locations temporally move over the subject's body parts in images while tack- ling the tasks of determining the subject's impression. We also investigated how the temporal changes of the gaze locations vary when different impression words are contained in the tasks. Existing analytical studies have not considered the tem- poral changes of the gaze locations, though they have revealed the spatial alignment of the subject's body parts that gather gaze locations. In our experiments, we gave observers several tasks to determine the subject's impressions in images and measured their gaze locations in a time series while viewing the images. We computed the distance from the observer's gaze location to the subject's body part location at each time and evaluated the difference of the distances between the tasks. We found that the observer's gaze locations move away from the face and move to the upper and lower body parts variously as time passes, though the gaze location is gathered on the face at the initial time. We also found the impression words with similar and dissimilar temporal changes of the gaze locations among the words contained in the tasks.

In the commercial printing machine envisioned in this study, an electronic document is printed, and its image is captured by a built-in reader to detect defective parts (stains and irregularities) in the printed document. Basically, the inspection is a comparison between the submitted image and the printed image. However, the printed image contains geometric deformations and errors in density and color during reading, and the submitted image and the printed image will not match even if the printing is appropriate. To solve this problem, we proposed a new method of comparative inspection, which we call "Comparison KIZKI Processing". In experiments, we used a test chart and confirmed that we could detect small stains, streaks, and irregularities.

In this study, light structured light-based tunnel 3D measurement instrument without using background texture of camera images is proposed. 3D point clouds of tunnel cross sections are actively measured by a camera and a ring laser in one shot. To integrate coordinates of the measured cross sections without using background texture of camera images, an additional camera is used in the proposed instrument. To estimate positions and orientations of the cross sections by acquired images from the additional camera, a probabilistic 2D-3D point matching algorithm is proposed. For accuracy validation, a 3D measurement experiment was done in a tunnel. The result shows that coordinates of all the cross sections were properly integrated and a dense 48 m-long tunnel point cloud was achieved.

In scientific publications, information is conveyed in the form of figure and table as well as text. Among the many fields of research, computer vision focuses on visual information like image and video. Therefore, figure and table are useful to convey information in computer vision paper such as image used in the experiment or output of the proposed method. In the field of computer vision, conference papers are important, which is different from journal publications considered important in other fields. In this work, we study the use of figures and tables in computer vision papers in conference proceedings. We utilize object detection and image recognition techniques to extract and label figures in papers. We conducted the experiments from five aspects including (1) comparison with other field, (2) comparison among different conferences in computer vision, (3) comparison with workshop papers, (4) temporal change, and (5) comparison among different research topics in computer vision. Thorough the experiments, we observed that the use of figure and table has been changing between 2013 and 2020. We also revealed that the tendency in the use of figure is different among topics even in computer vision papers.

We have developed an algorithm to accurately detect unclear defects in X-ray image inspection of thick welded parts under low contrast, and strong noise. Statistical Reach Features (SRF) and High-order Local Autocorrelation (HLAC) are used as noise-robust feature extraction methods. In order to deal with a small number of defect samples, pseudo-defect data with actual noise is used for machine learning. When the discriminator is optimized for zero missing, the over-detection is significantly reduced, and the method is ready for practical application.

In this paper, we tackle anomaly detection in cluttered wide-field images. Typically, the conventional approaches perform patch-wise anomaly detection by cropping the original image. However, it is difficult to detect anomalies correctly when the image has significant variations and local anomalous areas. Therefore, inspired by the human visual inspection, we propose a novel anomaly detection framework called Gaze-based Anomaly Detection (GAD). Our GAD learns a gaze map obtained from inspectors and utilizes the map to pay attention to the anomalous areas. The experiment showed that the proposed method allows us to detect abnormal samples without cropping and localization pre-processing and outperforms the conventional ones.

Large-scale image datasets are constructed with a significant amount of time and human effort, however, an image may have biases in the camera positions that render the target object. In this study, we propose a framework that automatically conducts 3D models, multi-view images, and even category definition. We automatically generate 3D models based on fractal geometry, which is the regularity behind natural phenomena and render images from multiple viewpoints. By following those generation processes, we can automatically construct a generally large-scale image dataset that takes into account the viewpoint position. The experimental results show that the classification accuracy is improved over the conventional baseline in the context of pre-training for image recognition tasks. We show that our proposed method provides an effective method for automatically constructing pre-training datasets for image recognition tasks.

When inspecting defects such as scratches due to image processing, if we can obtain an image before defect occurrence, the defect can be detected by simply comparing the image pair before and after defect occurrence. However, this idea is generally unrealistic because it is impossible to obtain an image before defect occurrence from an image after the defect occurred unless we go back in time. Therefore, we propose a method of training a generation-based model to detect scratches based on subtraction between an input image and output image during testing. We obtained a large amount of ideal image pairs before and after defect occurrence (i.e., image pairs in which only defect regions are different and the others are almost completely the same) using an image-capturing device. An image with a background texture of high reconstruction performance was generated with our method by training based on this dataset. Although we used image pairs of only an aluminum plate for training, ROC-AUC measure was 0.9973 for a copper plate and 0.9904 for a stainless steel plate. This shows that our method is also effective for background textures different from those used for training.

This paper reports on sign language recognition based on human body part tracking. Tracking-based sign language recognition has practical advantages, such as robustness against variations in clothes and scene backgrounds. However, there is still room for improving feature extraction in tracking-based sign language recognition. In this paper, a tracking-based continuous sign language word recognition method called Spatial-Temporal Graph Convolution-Transformer is presented. Spatial-temporal graph convolution is employed to improve framewise feature extraction using tracking points, while Transformer enables the model to recognize word sequences of arbitrary lengths. Besides the model design, the training strategy also has an impact on the recognition performance. Multi-task learning, which combines connectionist temporal classification and cross-entropy losses, is employed to train the proposed method in this study. This training strategy improved the recognition performance by a significant margin. The proposed method was evaluated statistically using a sign language video dataset consisting of 275 types of isolated words and 120 types of sentences. The evaluation results show that STGC-Transformer with multi-task learning achieved 12.14% and 2.07% word error rates for isolated words and sentences, respectively.