With the arrival of IoT society, the use of network cameras is spreading rapidly, and coupled with the increase in resolution of video image sensors, an explosive increase in data traffic over the network is expected. For advanced usage of high-definition images obtained from individual cameras, it is indispensable not to process all data in the cloud, but to process them at the edge of the network and aggregate necessary information. Therefore, OKI is developing an AI edge technology that combines embedded image recognition and advanced deep learning technologies to achieve high-speed and lightweight computation with high environmental resistance for detecting vehicles and people from high-definition images (4k/8k resolution) covering wide road areas.
OKI has developed a unique "Frequency Superimposition Deep Learning" technology that combines the frequency domain representation and the brightness intensity representation of input image. As a result, it was possible to achieve lightweight and accurate models. OKI's models perform well even in bad weathers such as rain and snow.
Frequency Superimposition Deep Learning has the following features.
When this technology was applied to vehicle recognition task, the recognition accuracy under various weather conditions increased from 98.0% to 99.8%, the processing speed was increased 16 times, and the memory size was reduced to 1/10 (*1). This technology can also be applied to object detection task, and improvement is expected in performance of vehicle and human detection.