スマートホールディングス List of Products and Services

AI image inspection is an appearance inspection that utilizes deep learning, building models by learning the characteristics of the inspection targets from data. It enables easier operation and higher precision inspections compared to conventional inspection machines. Additionally, by replacing existing inspections with AI appearance inspections, it reduces the human costs associated with inspections. 【Technical Details (Partial)】 ■ Fusion of AI and Robotics - Many engineers with expertise in robotics are on staff, enabling automation of inspection processes involving movement through the fusion with robotics. ■ Wide range of image recognition technologies applied - YOLO Series, Mobilenet-SSD, DETR, DeepLab, DCGAN, VAE, etc. *For more details, please download the PDF or feel free to contact us.

We analyze visual information using AI to support communication from patients, including those with disabilities in medical facilities and at home, to doctors, nurses, and staff. Improvement of accessibility and quality of life is possible. It reduces the burden on medical staff and realizes the efficiency of medical care! Additionally, we provide customized care and treatment plans tailored to the individual needs of patients. 【Technical Details】 ■ Achieves high-difficulty tasks such as lip reading while wearing masks, pupil detection, and fine facial expression recognition. ■ Utilizes advanced AI algorithms and computer vision technology to enable real-time image processing and interpretation. ■ Ensures usability and effectiveness in real-world scenarios based on feedback from patients and caregivers. *For more details, please download the PDF or feel free to contact us.

The development of a driver monitoring system (DMS) using a single camera enables the detection of driver actions and ensures a clear line of sight to the road. By combining a system module box installed in the vehicle with a single camera, it monitors the driver's face and surroundings! Additionally, it uses an object detection model to detect and classify driver actions and to identify the presence of objects that may obstruct the driver's view. 【Technical Details】 ■ Camera Integration ■ Driver Action Detection ■ Real-time Monitoring *For more details, please download the PDF or feel free to contact us.

We are developing automation systems tailored to customer needs using machine learning technologies, including voice recognition and synthesis, large-scale language models, and retrieval-augmented generation (RAG). By connecting to inventory and product databases, we are implementing a system that allows tasks such as ordering in restaurants to be completed via voice, thereby reducing labor costs! In the personal assistant role, we handle general assistant tasks such as conversation and schedule management. 【Usage Image】 ■ Ordering and Inventory Management in Restaurants - By implementing a system that allows tasks such as ordering in restaurants to be completed via voice, we reduce labor costs. ■ Personal Assistant - Customization of voice and character is possible based on individual preferences. *For more details, please download the PDF or feel free to contact us.

In voice synthesis, generating speech with a natural accent or adjusting the accent arbitrarily while maintaining naturalness is challenging. However, by utilizing machine learning techniques, achieving a natural accent is possible. Implementations such as Supervised Fine-tuning (SFT) using human feedback enable the generation of natural speech! Please feel free to contact us when you need assistance. 【Technical Details】 ■ Implementations such as Supervised Fine-tuning (SFT) using human feedback enable the generation of natural speech. ■ By redesigning the loss function and other innovations, we reduce distortion in speech generation at extreme high and low frequencies. *For more details, please download the PDF or feel free to contact us.

We provide support for human resource development in various ways that align with your company's needs. We have engineers with business experience and PhDs in fields such as image recognition, natural language processing, and reinforcement learning. Additionally, we can offer training support in your preferred format, including group training, on-the-job training (OJT), and one-on-one sessions! 【Our Strengths】 ■ Capable of handling advanced and diverse technologies ■ Support available in various formats *For more details, please download the PDF or feel free to contact us.

To shorten the training time of deep learning, parallel computation using multiple GPUs is necessary. By exploring methods for efficient parallel computation, effective learning can be achieved! Additionally, we will explore and implement efficient learning methods tailored to use cases, such as model parallelism, data parallelism, and pipeline parallelism, and conduct experiments. 【Technical Details】 ■Execution of parallel learning suited to use cases ■Resolution of processing allocation issues in GPU placement *For more details, please download the PDF or feel free to contact us.

There is a problem of insufficient GPU memory for training large deep learning networks, while GPUs are expensive compared to HDDs and DRAM, making expansion difficult. By utilizing other hardware, it is possible to train large networks using only a single GPU! 【Technical Details】 ■ Data transfer to HDD using CUDA Unified Memory - Users can transfer data from GPU memory to Host memory without being aware of the transfer. - A technology has been implemented to transfer data to HDD when Host memory is insufficient by extending the Nvidia Driver. ■ Analysis of computation graphs - Development of a technology that keeps only the necessary data in the GPU, moving others to Host memory or storage, and transferring them back to the GPU when computation is likely needed. *For more details, please download the PDF or feel free to contact us.

In the field of NLP, one approach to improving accuracy is the introduction of memory layers. However, storing a large amount of memory can lead to the problem of slow retrieval times. To solve this, by accelerating approximate nearest neighbor search, it is possible to enhance the network's accuracy! 【Technical Details】 ■ Acceleration of Faiss - Implement state-of-the-art algorithms and aim for acceleration by integrating them into the network. ■ Development of approximate nearest neighbor search algorithms - Further tune Faiss to aim for overall acceleration during training. *For more details, please download the PDF or feel free to contact us.

The analysis of products dealing with fluids, structures, electromagnetic fields, etc., through numerical simulations of partial differential equations faces a significant challenge due to the long computation times. By using machine learning techniques, it is possible to significantly reduce computation time! Additionally, by executing simulations that have traditionally been performed on CPUs on GPUs, a speedup of up to 100 times can be achieved in the best case. 【Technical Details】 ■ Development of Faiss's machine learning-friendly numerical simulation algorithm for acceleration - By appropriately setting several parameters of the numerical simulation algorithms that have been traditionally used, we have developed an algorithm that can accelerate without compromising the traditional accuracy. - Fast numerical simulations can be realized through parameter optimization using machine learning. ■ Execution of numerical simulations on GPUs - By executing simulations that have traditionally been performed on CPUs on GPUs, a speedup of up to 100 times can be achieved in the best case. *For more details, please download the PDF or feel free to contact us.

To accurately perform AR in an external environment, the creation of point clouds is necessary. However, since point clouds typically need to be generated solely from stereo images from AR Glass, there may be cases where the accuracy of the point clouds is low. Therefore, by leveraging existing computer graphics and AI, high-precision point cloud generation is possible! We generate high-resolution 3D point clouds with 95% accuracy to meet customer needs. Additionally, rapid data processing enables real-time 3D scene reconstruction. 【Technical Details】 ■ Epipolar geometry-based multi-view stereo algorithms, including geometric correspondence, camera geometry, and scene geometry 　・MVS ■ Depth estimation networks for accurate depth map generation 　・DenseDepth, etc. ■ Stereo matching networks for end-to-end disparity map prediction 　・PSMNet, GCNet, etc. *For more details, please download the PDF or feel free to contact us.

The restaurant's AI assistant understands Japanese sentences and analyzes customer emotions, appropriately animating 2D anime characters. It comprehends customer orders and analyzes emotions, animating 2D anime characters according to the customer's feelings (e.g., smiling, sad expressions). It achieves basic animations through phoneme extraction and advanced animations using diffusion models! (e.g., head movements, speed control) 【Technical Details (Partial)】 ■ Preprocessing - Uses morphological analysis tools (e.g., MeCab, Janome) to tokenize Japanese text - Removes unnecessary information (stop words, special characters) *For more details, please download the PDF or feel free to contact us.

Automation of object detection and processing through the integration of AI, computer vision, and robotics in the manufacturing industry! By automating tasks, we streamline repetitive operations, allowing personnel to focus on high-value activities. Additionally, we optimize the effective use of workflows and resources, enhancing overall productivity in food preparation tasks. 【Technical Details】 ■ AI Camera Module: Utilizes a pre-configured camera module equipped with Mobilenet-SSD to achieve real-time object detection. ■ Implementation of Computer Vision Technology: Adoption of computer vision technology for accurate 3D object position estimation (x, y, z coordinates). ■ Instructions for Robot Arm: Provides instructions for the autonomous grasping and manipulation of detected objects. *For more details, please download the PDF or feel free to contact us.

Detect specific signs and actions within videos, understand the context before and after specific frames, and conduct analysis on videos such as dashcam footage and sports matches. Using existing methods, we will conduct accuracy experiments on our custom dataset to create a model that can detect the expected signs! 【Use Cases】 ■ Building a system to check the situation just before an accident or traffic violation occurs - Conduct video analysis and analyze driver habits, etc. - Create training materials for drivers ■ Building a system to detect the actions of people in videos - A system that detects not only people but also their actions - A system that detects specified actions - Crime prevention system *For more details, please download the PDF or feel free to contact us.

We will conduct data analysis and develop training to win international sports competitions. Using data such as heart rate before the competition starts and image processing technology, we will analyze the differences between good and poor athlete form and utilize this in training. Additionally, we will conduct mental training linked to heart rate and other values to eliminate mistakes during matches! 【Technical Details (Excerpt)】 ■ Perform three-dimensional composition and evaluation using DLT method, epipolar geometry, and PnP problem - Obtain the two-dimensional information necessary for three-dimensional composition using methods such as DeepLabCut and manual techniques ■ Investigate whether variations in the method of obtaining two-dimensional coordinates affect the accuracy of three-dimensional composition - Examine the relationship between the number of cameras used for shooting and their accuracy ■ Conduct research on methods for three-dimensional composition in situations where setting markers is difficult - Investigate the feasibility of using drones We also support sports × IoT technologies, including Body Sharing. *For more details, please download the PDF or feel free to contact us.

Our company has expert personnel in AI development and has a wealth of experience in contract development for applications such as image recognition and chatbots. We can provide one-stop solutions from design and implementation to evaluation and the construction of evaluation environments. We introduce technical details addressing challenges such as AI image inspection in the manufacturing and medical industries, as well as recognition of voice, emotions, and facial expressions to aid in human resource development and the efficiency of meeting minutes. We also cover topics related to image data, language data, reinforcement learning, and support for AI personnel development. Please feel free to download and take a look. 【Examples of Problem Solving (Partial)】 ■ Images - AI image inspection in the manufacturing industry - AI image processing in the medical industry - Improvement of automotive driving safety through object detection ■ Language - Voice dialogue automation systems - Enhanced controllability of voice generation *For more details, please download the PDF or feel free to contact us.