Hilbert-encoded Cooperative Coevolutionary Genetic Algorithm for Few-shot Learning

In recent years, the advancement of artificial intelligence algorithms, particularly in the domain of semantic segmentation within computer vision, has been noteworthy. This progress is attributed to the availability of diverse large-scale datasets and the continual improvement of computational resources, with deep learning serving as a representative paradigm. Consequently, numerous sophisticated segmentation models have emerged for visual tasks. However, the efficacy of these models hinges on the availability of substantial annotated datasets for training. A notable challenge arises in scenarios where few-shot datasets are prevalent due to inherent limitations. The sparse data volume in these instances often stems from difficulties in obtaining a sufficiently large and high-quality dataset. The process of image annotation is resource-intensive, demanding considerable time and human effort. Addressing the consequential shortage of sample size in few-shot datasets becomes a paramount concern. This paper tackles the aforementioned issues related to few-shot datasets and introduces the Hilbert encoded cooperative coevolutionary genetic algorithm (HCCGA), a novel image enhancement algorithm designed for the automated generation of 3D-Lookup Tables (3D LUTs). HCCGA leverages 3D LUTs generated automatically through the genetic algorithm to map the original few-shot dataset, thereby obtaining a more extensive set of high-quality samples. The methodology employed in this research encompasses the generation and optimization of 3D LUTs using a divide-and-conquer strategy based on genetic algorithms. Additionally, it involves the mapping of new images using 3D LUTs in conjunction with sampling and interpolation. The evaluation of the proposed approach encompasses comparisons based on various image binary classification evaluation metrics applied to images before and after mapping through the results of model inference. Experimental results obtained using HCCGA on diverse public datasets, combined with state-of-the-art segmentation models, demonstrate the superiority of the optimized 3D-LUTs-mapped samples over their unmapped counterparts. This superiority is evident in terms of various standard evaluation metrics, including precision, recall, and F1score. Furthermore, the high-quality images generated by this method offer a viable solution to augment the original dataset, effectively addressing the significant inadequacy in the number of samples within few-shot datasets.

I. Introduction

The continual proliferation of large-scale datasets and the concurrent improvement in hardware performance have propelled the progress of deep learning in computer vision, yielding significant scientific and technological advancements. Researchers have developed a diverse array of vision models, encompassing tasks such as object detection [1], semantic segmentation [2]–[3], image classification [4] and instance segmentation [5]. These models find applications across various domains to address diverse needs.