In 2020, due to the impact of the COVID-19 pandemic, the misuse of antibiotics has become more severe. This not only results in ineffective treatments but also contributes to increased bacterial antibiotic resistance, leading to the emergence of superbugs. Therefore, this we focuses on developing a bacterial image analysis system applicable to Antimicrobial Susceptibility Testing. The system employs phase-contrast microscopy to monitor bacterial cultivation within a micro fluidic device. The research entails the creation of a tailor-made micro fluidic chip integrated with an imaging system, facilitating the individual observation of bacteria. The system offers the advantage of minimizing liquid usage. The primary focus of this research is on calculating bacterial quantities and determining antibiotic resistance based on surface morphology. Convolutional Neural Networks, including YOLOv7, YOLOv8, U-Net, U-Net++, and U-Net3+, are employed for bacterial counting and surface morphology recognition. Ultimately, the performance validation results indicate that U-Net3+ and YOLOv7 exhibit prominent performance in model verification, demonstrating effective bacterial recognition. Subsequently, this system aims to enhance the detection of bacterial antibiotic resistance, contributing substantively to single-bacteria/cell recognition in the field of biological applications.

With the booming development of autonomous vehicles (AV) in recent years, a vehicle needs to have the ability to detect changes in the environment in real-time. If the vehicle can be decelerated in advance according to the traffic signs, it can effectively reduce fuel consumption and improve overall comfort. This paper uses the Kaggle data set for training based on marking the common traffic signs in foreign countries, adds the local data set in Taiwan to the testing data set, and uses the You Only Look Once v4 (YOLOv4) neural network to detect the traffic signs in real time. The experimental results show that YOLOv4 still has a good generalization ability in the case of slight differences in different national sign types, and the mean Average Precision (mAP) can reach more than 87.6%.

This research is based on the framework of a fully automated smart fruit factory that builds a small and simple fruit grading device, and spots defects in the three fruit models of apples, lemons and oranges which were used as test target, and the entire process of detection is performed in a dark box. There is a ring-shaped LED light to regulate the light source inside the dark box. The fruits to be identified are moved into the dark box by a conveyor belt, an infrared sensor is used to judge whether the fruits are within the shooting range of the image capture area, and then the photos are sent to the SSD (Single Shot Multi Box Detector) neural network model to identify defects. This system screens the surface of apples, lemons and oranges for defects like damage, pest damage, dryness, bruises, etc. and removes them to preserve the fruits that are good in quality. It has been verified by several experiments that the identification accuracy rate can reach upto more than 95%.

This research focuses on integrating low-cost 2D-Lidar into a multi-split laser vaporization system that can incinerate wastes and general objects to improve the problem of manual alignment of the laser vaporization machine. In the experiment, we proposed a custom angle range and a distance threshold screening method for 2D-Lidar to distinguish whether there is an object in the scanning area and object locating. Through the law of cosine, the 2D-Lidar scanning angle is evenly distributed to the multi-split laser vaporization system. The object scanning method uses the Modbus TCP communication mode to enable the 2D-Lidar on the PC side to handshaking with the PLC on the central control side of the system. Scanning and size calculation of the object can obtain the position point of the object and hence improving the efficiency of the multi-split laser vaporization system. In addition, 2D lidar is readily available and inexpensive, in which making it more cost-effective than 3D-Lidar.

Eco-driving is an effective and immediate environmental protection and energy saving method. This research assists occupational driving license training to achieve eco-driving at two parts: 1. Combine g-sensor with on board diagnostics (OBD-II) and add parameters to improve the data analysis. 2. Through two kinds of neural network models, predict fuel consumption to analyze driving style, and provide reports to display evaluation and behavior suggestions. The experimental configuration designed in this research includes user interface, OBD-II system, neural network model, and is applied to public institutions to provide assistance. The results of this study show that the accuracy of predicting fuel consumption exceeds 97%, which verifies the practicability of the system. The system will also help extend other related applications, such as achieving a driving behavior model that compares energy saving and safety.