面向小样本图像分类的生成式数据扩增与APCRLNet融合方法

A Fusion Method of Generative Data Augmentation and APCRLNet for Few-Shot Image Classification

摘要:

小样本图像分类中存在样本不足、易过拟合及部署效率受限等问题,由此文章提出一种融合生成式数据扩增与APCRLNet的分类方法。采用改进GAN、泊松编辑和MS-Net生成扩增样本,并结合SSIM筛选与加权训练提升数据有效性；构建含非对称并行卷积与残差学习的APCRLNet,并通过通道剪枝和8 bit量化实现轻量化。实验表明,本文方法Top-1准确率达95.2%,较对比方法分别提高6.7、8.9和10.5个百分点,参数量减少42%,推理时间缩短45%,实现了精度与效率的协同提升。

Abstract:

To address insufficient samples, overfitting, and limited deployment efficiency in few-shot image classification, a method integrating generative data augmentation and APCRLNet is proposed. Augmented samples are generated by an improved GAN, Poisson editing, and MS-Net, and optimized by SSIM-based screening and weighted training. APCRLNet combines asymmetric parallel convolution and residual learning, while channel pruning and 8-bit quantization are used for lightweight deployment. Experiments show that the proposed method achieves a Top-1 accuracy of 95.2%, improving by 6.7, 8.9, and 10.5 percentage points over comparative method, respectively, while reducing parameters by 42% and shortening inference time by 45%.