"Faster and More Precise: Researcher Improves Performance of Image Recognition Neural Network"
There have been advancements in Machine Learning (ML) regarding image recognition as this technology can now identify objects in photographs and videos. The adoption and implementation of image recognition continue to grow. However, such systems still call for improvements. Andrey Savchenko, a Professor at HSE University, developed an image recognition algorithm that functions 40 percent faster than analogues. It has been demonstrated to be capable of speeding up the real-time processing of video-based image recognition systems. Convolutional Neural Networks (CNNs) include a sequence of convolutional layers. They are widely used in computer vision. Savchenko was able to speed up the work of a pre-trained CNN using arbitrary architecture, containing 90 to 780 layers. This resulted in a 40 percent increase in recognition speed while at the same time controlling accuracy loss to no more than 0.5 to 1 percent. He used statistical methods like sequential analysis and multiple hypothesis testing. High accuracy is essential for image recognition systems. An incorrect decision made by a face recognition system can lead to someone from the outside gaining access to confidential information or the user being denied access repeatedly due to the neural network's inability to identify them. Sometimes speed can be sacrificed, but it is important in the application of video surveillance systems where there is a desire to make decisions in real-time. Professor Savchenko emphasizes the need to recognize an object in a video quickly without losing accuracy. This article continues to discuss the image recognition algorithm developed by Professor Savchenko that can speed up the real-time processing of video-based image recognition systems while controlling the loss in accuracy.