Project

This technology features "zero-contact" detection, we observe the micro-vibration of the head caused by the contraction of the heart, and develop a zero-touch facial heart rate and breathe detection technology in conjunction with the camera. The technology can accurately measure heart rate and respiration in real-time, thereby reducing the risk of infection.

The chest X-ray is a radiological clinical assessment tool used to detect different types of lung diseases, such as lung tumors. We use SDFN (Segmentation-based Deep Fusion Networks) and SE (Squeeze and Excitation) Blocks for model training, using a combination of whole and cropped lung X-ray images, which also assists in improving the model’s attention, avoiding issues that could be introduced by image misalignment and unwanted objects, as well as the loss of small targets after image resizes. Two CNNs are used for feature extraction, with the extracted features being stitched together to form the final output that is used to determine whether lung tumors are present. Unlike previous methods of identifying lesion hotspots from X-ray images, we use SEG-GRAD-CAM to generate heatmaps of the lung tumors for localization. From experimental results, we achieved a 98.51% accuracy and 99.01% sensitivity in classifying chest X-ray images with and without tumors. The method can reduce errors caused by differing judgments between radiologists, and assist them in making medical decisions.

A dynamic random access memory (DRAM) module contains a lot of electronic components. DRAM modules should be carefully inspected before leaving the factory. Existing automatic optical inspection (AOI) machines have poor detection performance and often misjudge normal items as abnormal items. As a result, the workload of following manual inspection is highly increased, which not only wastes time, but also increases the risk of misjudgment caused by human eyes fatigue. In order to reduce the misjudgment rate, we develop a novel generative adversarial network (GAN) deep learning method for DRAM module appearance defects detection. Experimental results showed that the defect detection accuracy rate is 99% and the defect missed detection rate is less than 1%, which effectively solves the problem that the traditional AOI equipment has a high misjudgment rate and requires a large amount of manual re-inspection.

Our technology combines infant crying recognition, infant mouth and nose occlusion detection, and heart and respiration rate detection. We use incremental learning with deep neural networks to overcome the limits of a traditional support vector machine (SVM) classifier for cry recognition. In our heart and respiration rate detection, we addressed problems associated with conventional no-contact technology, such as inaccuracies due to ambient light, improving the detection accuracy. The infant crying recognition technology has been commercialized successfully.

The learning observation system captures the learner’s face and movements and analyzes the expression information (including happy, sad, and expressionless) and specific actions (Including raising left hand, raising right hand, raising hands, and getting down), assisting teachers to understand the emotional changes of students in class and student’s reaction, can find the students that need to be paid attention to, and adjust teaching methods or communicate effectively in real-time.

The important issue of industrial safety accident prevention is to avoid the employees who do not follow the safety procedures and seek the convenience to cause industrial safety incidents. Therefore, a set of AI Monitoring Systems for Safe Behavior of Entrances/Exits on Production Line was developed, using pedestrian tracking and behavior recognition technology to timely monitor whether employees follow safety standards and improve the safety of employees’ operations.

Although digital shearography can show subtle defects that cannot be observed with the naked eye, it is still up to the site personnel to determine whether it is a defect, and the judgment standards may vary due to different experiences. Our team proposed the incremental YOLO architecture to greatly increase the detection rate. The bubble defect detection rate is about 98%, and it only takes 0.076 seconds to judge a single image. The detection speed is very fast, which can help companies achieve semi-automatic detection processes and reduce the detection of human resources.

The contributions are as follows. (1) Realization of automatic thyroid gland segmentation in various medical images, (2) Segmentation and classification of thyroid nodules in ultrasound images, (3) Volume estimation of the thyroid gland in ultrasound images, (4) Segmentation of thyroid tumors (5) automatic detection of the Graves disease in the thyroid gland.