Wistron has been granted a patent for an object identification method that involves generating tracking and adversarial samples, training a teacher model, and initializing a student model. The student model adjusts parameters based on the teacher model and adversarial sample, and is considered trained when the vector difference between its output and the teacher model’s output is below a learning threshold. The student model is then extracted as an object identification model. GlobalData’s report on Wistron gives a 360-degree view of the company including its patenting strategy. Buy the report here.
According to GlobalData’s company profile on Wistron, Device power optimization was a key innovation area identified from patents. Wistron's grant share as of September 2023 was 76%. Grant share is based on the ratio of number of grants to total number of patents.
Object identification method and device using teacher-student model training

A recently granted patent (Publication Number: US11776292B2) describes an object identification device and method that utilizes deep learning models to improve object identification accuracy. The device includes a processor and a storage device, with the processor being configured to access programs stored in the storage device to implement various modules.
The device includes a pre-processing module that generates a tracking sample and an adversarial sample. A teacher model training module uses the tracking sample to train a teacher model, while a student model training module initializes a student model based on the teacher model. The student model adjusts its parameters according to the teacher model and the adversarial sample. Once the vector difference between the output result of the student model and the output result of the teacher model is less than a learning threshold, the student model is considered to have completed training and is extracted as the object identification model.
The deep learning model configured by the student model has a lower number of convolutional layers, neurons, and weight parameters compared to the teacher model. This allows for more efficient and streamlined object identification.
The pre-processing module also receives a manually labeled sample. When the vector difference between the output result of the student model and the output result of the teacher model is lower than the learning threshold, the teacher model training module inputs the manually labeled sample into the teacher model to generate an advanced teacher model. This advanced teacher model is further trained by inputting the manually labeled sample and comparing the post-training tensor output to the manually labeled sample.
The student model training module adjusts the parameters of the student model according to the advanced teacher model. Once the vector difference between the output result of the student model and the output result of the advanced teacher model is lower than the learning threshold, the student model is considered to have completed training and is extracted as the object identification model.
The pre-processing module tracks a frame-selected object using an optical flow algorithm to generate the tracking sample. It can also add noise to the tracking sample or input it into a generative adversarial network (GAN) to generate the adversarial sample. The noise map added to the tracking sample includes images of different object types, enhancing the robustness of the object identification process.
In summary, this patent describes an object identification device and method that utilizes deep learning models, pre-processing techniques, and advanced training methods to improve object identification accuracy. The device and method offer a streamlined and efficient approach to object identification, with the student model being extracted as the final object identification model once it has completed training.
To know more about GlobalData’s detailed insights on Wistron, buy the report here.
Data Insights
From
The gold standard of business intelligence.
Blending expert knowledge with cutting-edge technology, GlobalData’s unrivalled proprietary data will enable you to decode what’s happening in your market. You can make better informed decisions and gain a future-proof advantage over your competitors.