Bit Width Optimization of Swish Function in YOLOv8 Object Detection Model

Recently, neural-network-based algorithms are showing excellent performance in computer vision tasks including classification, object detection, and segmentation. A neural network model for a vision task usually consists of various convolutional layers and fully connected layers, and activation functions are inserted between the layers. The swish function is being widely used as an activation function, but the quantization of the swish function has not been analyzed well. This paper performed experiments with various bit widths on the swish function of YOLOv8n object detection model. The results showed 8-bit quantization with the integer part of 3-bits is the best choice preserving the original detection capability.