Deep Neural Network (DNN) is the core area of Artificial Intelligence (AI), which deals with algorithms that mechanically mimic the biological structure and function of brains. The deep learning model consists of a number of neurons, activation functions, optimization algorithms, data addition tools, etc.
Deep learning based on artificial neural networks is well suited for modeling, classifying and recognizing complex data such as images, speeches, texts, language translation, etc. However, the use of deep neural networks requires a large amount of training data. In order to identify personal information, for example, you need to record a large number of pictures and voices of each person for training. The large amount of training data leads to a huge amount of computation, and therefore, they are distributed to the computers connected to the Internet worldwide, or high-performance computers equipped with GPUs are used.
Kim Sun Il, a researcher at the Faculty of Metal Engineering, has established a detection system to increase the detection speed and rate of amorphous body detection, using deep neural network that is widely used worldwide, and evaluated its performance. He configured the deep neural network suitably for amorphous body detection based on YOLOv4 that provides high speed and detection rate for object detection.
First, he modified the K-means clustering method used in the standard YOLOv4 to fit the database used to increase the detection rate. Then, he modified the structure of the model to increase the detection speed by reducing the number of overlapping layers to 16 and the number of parameters to 17 059 472.
With the proposed deep neural network, the detection rate is 98% and the time of processing the image of a frame is 0.04s, which demonstrates its high detection accuracy and real-time performance.
© 2021 Kim Chaek University of Technology