新型快速機械/電腦視覺關鍵技術開發與智能自動化多目標生物特徵辨識與追蹤模組之整合應用

Based on the support of MOST projects, we have developed a novel semiunsupervised artificial intelligence technology for computer vision feature recognition and automated image feature enhancement technique. In subsequent research, we successfully developed theories applicable to various data structures based on DFT and preliminary research in computer vision and edge computing. Thus, we call this technique fast data density functional transformation (fDDFT). Since fDDFT is regulated under the laws of physics, the data structure of Hilbert space can be regarded as quasi-physical particles, which makes fDDFT have a behavior of spontaneous self-attention and can learn independently. In the previous research, the spontaneous self-attention learning module behaviors as the core of unsupervised segmentation technology for brain tumor MRI and multi-target images. Therefore, in this three-year proposal, we will further develop and complete the theoretical framework. Then we will also apply the spontaneous self-attention learning module to the problems, such as multigrain tumor segmentation, color medical images segmentation, real-time identification and tracking of multi-object images, and deep modeling and 3D reconstruction of multi-modal high-dimensional medical images. Additionally, the computational complexity is low because input images would automatically trigger the self-learning module. In terms of mathematical structure, this module is only a well-defined reciprocal convolutional kernel that acts on the input image, so it is very suitable for edge computing applications. In the previous research, we have confirmed the feasibility of this module on edge computing. Therefore, in the later period of this three-year proposal, we will continue to apply this module to key biomedical issues such as identification, tracking, and segmentation of biomedical images. We will also establish the algorithm in a low-cost microprocessor of smart sensor modules for biometric identification. We look forward to using this three-year proposal to promote this fDDFT method with spontaneous self-attention and self-learning modules. Furthermore, we also expect that this method can change the infrastructure of contemporary deep learning and computer/machine vision. Hopefully, it can become the main core of artificial intelligence.