Evolutionary Simulated Annealing for Transfer Learning Optimization in Plant-Species Identification Domain

Abstract

The reuse of the pre-trained deep neural network models has been found successful in improving the classification accuracy for the plant species identification task. However, most of these models have a large number of parameters, and layers and take more storage space which makes them difficult to deploy on embedded or mobile devices for real-time classification. Optimization techniques, such as Simulated Annealing (SA), can help to reduce the number of parameters and the size of these models. However, SA can easily get trapped into local optima when dealing with such complex problems. To solve this problem, we propose a new technique, namely Evolutionary Simulated Annealing (EvoSA), which optimizes the process of transfer learning for the plant-species identification task. We incorporate the genetic operators (e.g., mutation and recombination) on SA to avoid the local optima problem. The technique was tested using the MNetV3-Small as a pre-trained model due to its efficiency on mobile for two plant species data sets (MalayaKew and UBD botanical garden). As compared to the standard SA and Bayesian Optimization techniques, the EvoSA provides the least cost value with a similar number of objective evaluations. Moreover, the EvoSA produces approximately 14x and 6x less cost compared to SA for MalayaKew and UBD botanical data sets, respectively. The results show that the EvoSA can generate solutions with higher test accuracy than typical transfer learning with a competitive number of parameters.