Target recognition algorithm based on improved depth separable convolution

摘要

Abstract Realizing accurate target recognition in the fields of drones, vehicle-mounted systems, robots, etc. has increasingly become a hot research direction for researchers. Aiming at the problem of using deep learning algorithms to identify targets on mobile devices with limited size, the network model has a large number of parameters resulting in a large calculation cost, a complicated recognition process, and a large number of iterations resulting in a long recognition time. This paper proposes an improved deep separable convolution algorithm, which consists of deep convolution and point-by-point convolution, which is based on convolution kernels for step-by-step filtering and combination. After deep convolution, batchnorm is added, and the activation function uses a linear function. Add batchnorm and nonlinear activation function ReLU after 1×1 convolution to ensure the accuracy of target detection. This algorithm reduces the amount of parameters while ensuring recognition accuracy, reduces the amount of recognition calculations, and reduces the weight of the network model. It provides a network foundation for implementing image recognition on devices with limited size.