EasyDAM_V4: Guided-GAN-based cross-species data labeling for fruit detection with significant shape difference.

Area of Science:

• Agricultural Technology
• Computer Vision
• Machine Learning

Background:

• Smart orchards rely on high-performance fruit detection, which requires large labeled datasets.
• Current methods for automatic labeling struggle with fruits exhibiting significant shape variations.
• Manual data labeling is costly and time-consuming, hindering AI development in agriculture.

Purpose of the Study:

• To propose an improved automatic fruit labeling method, EasyDAM_V4, to address limitations in cross-domain fruit image translation.
• To effectively reduce domain differences by translating phenotypic features like shape, texture, and color.
• To enhance the applicability of automatic labeling for fruits with substantial shape variance.

Main Methods:

• Utilizing an improved fruit automatic labeling method, EasyDAM_V4.
• Introducing the Across-CycleGAN fruit translation model for spanning translation between source and target fruit images.
• Validating the method on pear fruit (source domain) and pitaya, eggplant, and cucumber (target domains) with large phenotypic differences.

Main Results:

• EasyDAM_V4 demonstrated substantial cross-fruit shape translation capabilities.
• Achieved average labeling accuracies of 87.8% for pitaya, 87.0% for eggplant, and 80.7% for cucumber.
• Effectively reduced domain differences between fruits with significant shape variance.

Conclusions:

• The EasyDAM_V4 method significantly improves automatic fruit labeling, even with considerable shape differences between domains.
• This research enhances the practical application of AI in smart orchards by reducing data labeling costs and improving model training.
• The Across-CycleGAN model offers a robust solution for cross-domain image translation in agricultural computer vision tasks.