Particle Swarm Optimization-Based Transfer Learning for Image Segmentation and Classification

Appropriate model hyperparameters, transfer learning strategies, and loss balancing are still challenging to discover, while deep learning has achieved remarkable progress in picture segmentation and classification. An integrated encoder-decoder network is improved in this transfer learning system with the application of Particle Swarm Optimization (PSO). To get hierarchical data, a pretrained encoder is used, and lightweight segmentation and classification heads learn two tasks at once. By default, PSO, a universal meta-optimizer, changes learning rates, layer-freezing masks, batch sizes, and loss-weighting coefficients. Manual searches are accelerated in this way. With this hybrid approach, we may improve the accuracy-efficiency trade-off by combining global exploration with gradient descent and Particle Swarm Optimization (PSO). Cityscapes, PASCAL VOC, and CIFAR simulations demonstrate that swarm-based meta-learning has the potential to increase the flexibility of transfer learning, leading to more efficient and less expensive models for picture processing.