A data augmentation method for computer vision task with feature conversion between class

doi:10.1016/j.compag.2025.109909

Return

A data augmentation method for computer vision task with feature conversion between class

2025-04-01

PRE

Jiewen Lin

Gui Hu

Jian Chen *

DOI：10.1016/j.compag.2025.109909

Original

Original request for help

Save

Abstract

En 中文

Agricultural samples are unbalanced, complex, and scarce, which is the main factor restricting the popularization and application of agricultural computer vision. This paper proposes a feature conversion between classes method for data augmentation of computer vision tasks. We make contributions in the following three aspects: 1) Proposing an optimization method of attention mechanism to optimize the generator of CycleGAN. Through the module: efficient convolutional block attention model (ECBAM), the generator network structure of CycleGAN is improved to learn the feature transformation from healthy leaves to fake diseased leaves. 2) An label assignment method based on proportionally assigned receptive field is proposed to realize the label replacement from healthy leaves to fake diseased leaves. 3) Enhanced the original data by a factor of n x oversampling. The experimental results show that the improved CycleGAN proposed in this paper can effectively generate fake diseased leaves, the Inception Score (IS) is 2.3 +/- 0.14, the Frechet Inception Distance (FID) is 41.49, and the Kernel Inception Distance (KID) is 0.025. We have verified the feasibility of the method for classification, object detection, and semantic segmentation tasks. When using the improved CycleGAN for data augmentation, the accuracy of ResNet152 has been improved by 1.71 %. We further verified the effectiveness of improved CycleGAN and reactive field object assignment(RFOA) methods for data augmentation. By testing in the object detection task, when t = 0.75, and n = 1, the mAP reaches 78.97 %. By testing in a semantic segmentation task, when t = 0.50&0.75, and n = 2, the mIOU reaches 81.41 %.

Keywords:

Data Augmentation

Nutrient Deficiency

GAN

Attention Mechanism

Receptive Field