Recently, genomic information has been used for plant breeding. Thanks to DNA sequencers, genomic sequencing became easier, but it is not revealed that which genes are related to which traits. In order to use genomic information to plant breeding, we need to search the correspondence between genes and traits. In the situation, phenotyping, which grows various plants and measures traits exhaustively, becomes popular. Image processing and pattern recognition are commonly used for phenotyping because they realize mondestructive measurements and can acquire both 3D structure and color information. In this article, I will introduce the problems of plant measurements using images, and trait measurement methods using image processing and pattern recognition.

This article presents an automatic matching method for sprint performance videos based on the distance between sprinterʼs body joints. This technique makes it easy to compare a sprint form (e.g. one to be improved) with another (e.g. the ideal one), and consequently, enables us to efficiently find problems on the sprint form. Our method inputs performance video pairs captured at two different points (e.g. 50 m vs. 90 m) in a 100 m sprint, and applies sprinter detection and pose estimation. Then, our method matches the video pairs within a dynamic time warping framework based on the sprint form similarity between video frames. Experimental results showed the effectiveness of the framework of our method.

Deep-learning-based computer vision technologies have been commoditized owing to the synergy among various software infrastructures that enhances experimentʼs reproducibility. As applications of such a "commoditized" object detection method, this paper introduces two case studies, spatial and temporal analyses of task-execution pattern through an observation of object arrangement on a working table. In the spatial analysis, we extracted concept models given by each worker as an area segmentation on the flat workspace surface and role-assignment to each segmented area. This was done by an approach to treat the detected objects as if they were words in a document. In the temporal analysis, we found some specific patterns that relates to workerʼs strategy in cooking, by visualizing time series variation in the number of detected objects in a specific area on the working table.

With the development of machine learning technologies and the spread of mobile terminals, cloud-based image recognition services are getting popular in recent years. However, these services might suffer from a new type of attacks called retraining attack (RA), in which an attacker sends a lot of images to a recognition server and receives their recognition results to train a recognizer that mimics the serverʼs recognizer. We refer to the recognizers trained by RA as recognizer clones and aim to develop a defending method against them in our ongoing research project, whose current status is reported in this paper. Specifically, we consider the following two approaches: One is a method for preventing attackers from training recognizer clones by intentional misrecognition, where the server intentionally misrecognizes the images sent from the attackers. The other is a method for detecting already trained recognizer clones by checking the characteristics of their recognition results. While these two methods are still under development, we obtained some interesting knowledge through our experimental results.

Magnetospinography system is a device that measures non-invasively the magnetic field from the spinal cord, estimates the reconstructed current distribution around the spinal cord from the magnetic field and visualizes it. To estimate this distribution, itʼs necessary to extract the region of the spinal cord in a low dose plain X-ray image. However, since this region is extracted manually, the processing up to the estimation of this distribution isnʼt automated. As a result, these lead to high inspection cost. Therefore, we aim to automatically extract this region from a low dose plain X-ray image and conducting our research in three steps: i) detection and labeling of the vertebral body part, ii) contour extraction of the spinal vertebral body part, iii) extraction of this region. In this paper, as a first step, we propose the method to automatically detect and labeling lumbar vertebral body, intervertebral disc and sacrum from a lumbar lowdose plain X-ray image using CNN. Itʼs possible to robustly detect each part in high noise and low contrast images by devising CNN pre-processing and post-processing. The performance of the proposed method is evaluated using 60 plain X-ray images, and the results show the effectiveness of our approach with labeling precision reaching 93.3％.

The Demons algorithm is known as one of the classical methods for deformable image registration based on signal value difference and signal gradient, and its use is basically limited to for monomodal and scalar images. In this study, we show that the signal value difference and signal gradient used in the algorithm can be replaced with generalized distance between local features of two images and its gradient, such as color distances of color images. Through the experiments by using volume data sets of synthetic color images and diffusion tensor colormap images, we validate the proposed method and discussion its characteristics.

Diffusion-weighted MRI (DWI) has become an indispensable tool in clinical diagnostic imaging. This column will explain the merit of DWI in clinics by focusing on to some of the representing imaging and analyzing methods that have drove DWI to todayʼs position.