This paper describes training data to be applied to a Deep Neural Network (DNN) to automate the visual inspection of large die-cast parts with complex 3D shapes. In visual inspection systems, an alignment of a target workpiece is an important issue. If the workpiece is planar, it can be aligned by geometric transformation using image processing. However, when the workpiece is three-dimensional, alignment by image processing is essentially impossible. When the workpiece is large and has a complex three-dimensional shape, the difference in appearance due to misalignment is significant. In this study, we propose a method to reproduce differences in appearance images caused by misalignment by controlling the inspection system. By using the images reproduced by the proposed method as training data, the inspection accuracy of the appearance inspection DNN is improved. Experiments using actual inspection images confirmed the effectiveness of the proposed method.
We propose a fast and reliable method for detecting objects using color information. The probability of the occurrence of a hue pattern in a template image is calculated for each two-pixel pair, and only those pixel pairs with extremely low probability are carefully selected for matching. Since such pixels have high distinctiveness, it can be used for highly reliable matching without being affected by ambient disturbances, and since only a small number of pixels are used, it is also fast. In real image experiments, we achieved a recognition success rate of 93% and a processing time of 47 msec by using only 0.5% (143 pixels) of the template image. When 0.1% (29 pixels) of the template image was used, the recognition success rate was 7 msec, confirming that practical matching with a good balance between reliability and speed was achieved.
Crop detection is important for automatic weeding. A previous study achieved crop detection by detecting plant blobs in crop rows and by classifying them as crop or weed using k-means. However, this method was not robust to the amount and shape variation of the weed. In this paper, we propose a crop detection method that uses the uniform spacing in crop positions to address the problem encountered by the previous method. The classification results of the previous method are updated such that plants with uniform spacings are classified as crop. Experiments with artificial and real weeds validate the effectiveness of the proposed method.
Since the temperature distribution of the face varies among individuals, there is a possibility that it can be used for personal authentication. This paper proposes a method of using the high temperature region of the face, which is stable during personal authentication and resistant to changes in room temperature. The method obtained a 87% recognition rate in individual identification experiments under a variety of room temperature. In addition, this paper experimentally explains the method is stability against internal and external factors such as mental stress load and direct wind. Furthermore, combining this method with conventional face recognition methods using visible images, it will be possible to improve the authentication rate and avoid various risks.
The spline filter (SF) used the inverse matrix can measure the roughness without wasting the data at both ends of the input data. But when input data include outliers, the output of the SF greatly fluctuates. Therefore, “Fast M-estimation Spline filter (FMSF)” was proposed. In order to apply the SF to the fast M-estimation method, convolution-type SF is used instead of inverse matrix-type SF. FMSF performs robustly to input data including outliers, and gives the same output as that of convolution-type SF to input data without outliers. However, the convolution-type SF is made from an approximate formula. It needs a larger filter to improve accuracy of SF. This require a huge computational cost for convolution. There is a trade-off between approximation error and computational cost. To solve these problems, we propose “frequency-domain-type FMSF.” This allows the use of the SF transmission characteristics function instead of the convolution-type SF. The SF transmission characteristics function has no approximation error and exactly matches the SF output. Furthermore, the FMSF requires O(N2) computational cost. However, by using Chirp-Z transform to discrete Fourier transform, the proposed method can reduce the computational cost to O(N logN).
This study addresses a problem in the sensory inspection of wood products. Since wooden products have large variations in color and pattern, it is challenging to learn the distribution of good products using conventional anomaly detection methods. In addition, the boundary between good and bad products in the sensory inspection of wooden products is ambiguous and difficult to discriminate correctly since the decision criteria depend heavily on the inspector's sensitivity. Therefore, we propose a novel method to determine the decision boundary according to sensibility obtained from GAN's latent space traversal and sensibility surveys. We found that the proposed method is more effective than conventional methods for detecting abnormalities in the sensory inspection of wood grain images.