AVNPSO: Hyperparameter Optimization of Encoder-Decoder Networks for Image Segmentation

Image segmentation is crucial in many fields, but existing image segmentation models based on encoder-decoder networks are constrained by manual parameter tuning and the limited applicability of the selected parameters. To solve this problem, an adaptive particle swarm optimization algorithm based on Von Neumann structure (AVNPSO) is proposed for structural hyperparameters and learning hyperparameters optimization of encoder-decoder networks. The AVNPSO algorithm organizes the particle connections through the Von Neumann topology, prompts the particle information interaction and state update by the fitness value and adaptive mechanism of the network computation, and completes the optimization process of network structural hyperparameters and learning hyperparameters through the interaction between the network and particles. Furthermore, the AVNPSO algorithm adaptively updates the inertia weights and learning factors of the particle swarm algorithm according to the population entropy and clustering index, so that the algorithm can update its behavior according to dynamic changes of the population and better adapt to the needs of different search stages. The Von Neumann topology defines the neighborhood relationship, thus facilitating the exchange of particle information. This allows the particle to better balance the global and local search during the search process, effectively prevents falling into local optimal solutions, and improves the search efficiency. After experimental verification on two remote sensing image datasets, the optimized encoder-decoder network based on the AVNPSO algorithm significantly outperforms conventional methods, with an improvement in precision of more than 6%, demonstrating the effectiveness of adaptive optimization of network parameters.