Design of nonlinear segmentation activation functions for object detection

The existing activation functions ReLU, Tanh, and Mish have problems such as "neuronal death", offset, poor robustness. Aiming at these problems, the XExp function is proposed by combining advantages of Swish, functions, problem negative half-axis neuronal death optimized using nonlinearity non-RELU family non-zero characteristics soft saturation semi-axis retained. By designing position origin function, positive offset in Swish are solved. In terms convergence speed, MNIST dataset achieved 93.87% training accuracy during first batch on newly which was more than 85% higher speed compared with Relu function; model stability, can still achieve 98.05% when number convolutional layers increased to 25 layers. two data sets CIFAR-10 CIFAR-100 verify their versatility practicality field object detection.