Vehicle Feature Recognition via a Convolutional Neural Network with an Improved Bird Swarm Algorithm

Abstract

Accurate vehicle feature recognition is an important element in traffic intelligence systems. To address the problems of slow convergence and weak generalization ability in using convolutional neural networks to improve vehicle feature recognition, we propose an improved bird swarm algorithm to optimize convolutional neural networks (IBSA-CNNs) for vehicle recognition strategies. First, we use the center of gravity backward learning strategy and similarity- and aggregation-based optimization strategy in population initialization and foraging behavior, respectively, to improve the algorithm performance and avoid falling into a local optimum. Second, the improved bird swarm algorithm is used to optimize the weights of the convolutional and pooling layers of the convolutional neural network to improve the neural network performance. Finally, we tested the performance of the improved bird swarm algorithm in simulation experiments using benchmark functions. The recognition performance of IBSA-CNN was tested by the UCI dataset, and in the traffic vehicle dataset BIT-Vehicle, it improved 4.9% and 6.8% compared with R-CNN and CNN, respectively, indicating that IBSA-CNN has better vehicle feature recognition.