Abstract: The few samples generation technology relies solely on scarce and limited target samples to generate images that are both fake and diverse,which can build reliable datasets for downstream target recognition tasks.In this paper,we propose a few sample generation model based on weight modulation,which can obtain images with the same content and diverse feature representations as the target samples under the condition of only inputting three target images.Specifically,we have carefully designed the encoder and decoder in the generator,using a C2F structure with better gradient flow to build a pyramid network architecture,maximizing the restoration of the original features of the image at different levels.We adopt a feature fusion method based on attention mechanism and introduced feature style latent codes to control the quality of feature fusion.Among them,the style latent code uses a weight scaling strategy,effectively eliminating generated artifacts and making the generated images more realistic.At the same time,we also use an optimized feature length detection algorithm to detect the proximity of important information in the source and target domains.This technique enables the model to better transfer the prior information obtained through pre-training in the source domain to the target domain.For the task of generating flame image samples,we provide qualitative and quantitative comparison results.The proposed model can effectively improve the flame target recognition performance under the YOLOv8 algorithm and substantially enhance the data augmentation effect.

Key words: Few samples generation, Flame dataset, Feature fusion module, Transfer learning, Pre-training, Weight modulation, Target recognition, Style latent code