When a Convolutional Neural Network (CNN) classifier with the input of a small patch image evaluates the state of rust on weathering steel, the evaluation result depends on where the patch image is cut out. In this paper, we propose two types of methods to improve rust state evaluation using CNN. Both types employ multiple CNN classifiers based on bagging, which is one of ensemble learning. The final judgment result is derived by integrating the judgment results of multiple CNN classifiers. The first method is a method for improving the estimation accuracy of each patch image, and the integration of judgment utilizes majority voting and judgment probability. The second method is an evaluation method for a large tape image. Its evaluation result is derived by using a plurality of patch images cut out from the large tape image as input of CNN classifiers. Numerical experiments show the effectiveness of the proposed method.
Road administrators inspect roads daily. However, the number of facilities to be inspected is enormous, and local governments are seriously short of workers. The authors have focused on drive recorders, which are inexpensive and easy to install in patrol vehicles and have been conducting research to improve the efficiency of daily inspections using recognition technology of road objects by deep learning. In this study, we clarify the possibility of constructing a recognition model that extracts only the images of road objects suitable for inspection from video images and the optimal combination of road objects to be trained. From the results of the evaluation experiments, it was found that it is necessary to consider “only images suitable for inspection should be used as training data” and “only similar road features should be combined for training” when creating the recognition model.
In this paper, we measure the driver’s hand movement by using the wrist-worn acceleration sensors, and the driver activity is estimated by using the combination of the sliding-window method and k-nearest neighbors method. In order to evaluate the accuracy of driver activity estimation, we performed the experiment by using the driving simulator, and the data of 10 participants were collected. We show both the activity that can be detected by the wrist-worn acceleration sensors and its estimation accuracy.
The application of curriculum learning in reinforcement learning is expected to improve the learning efficiency and performance of autonomous agents by affecting their behavior acquisition. However, there are many challenges in generating the curriculum necessary for curriculum learning, as it is a prerequisite for having expert knowledge and deep prior understanding. In addition, curriculum generation methods are often specialized to the target task and lack versatility. In this study, new method is anticipated to solve the above problem by automatically generating a curriculum using clustering based on expert trajectories (state history). In the experiments, we compared the learning efficiency of normal agents and agents using the proposed method, and confirmed that the proposed method improved the performance in all aspects.