EasyDAM_V3 unveils next-gen automatic fruit labeling

In the dynamic field of agricultural AI, deep learning-based fruit detection has gained prominence, especially in smart orchards. However, these techniques are highly dependent on large, manually labeled datasets, a process that is time consuming and labor intensive.

The previous work introduced a generative adversarial network (GAN) method, EasyDAM, to reduce labeling costs by simulating fruit images. Nevertheless, this approach faces challenges: first, it lacks adaptability across diverse fruit species, leading to performance fluctuations in different orchard environments.

Second, while it reduces labor in the target domain, it still requires manual labeling in the source domain, and does not fully eliminate manual processes. There is a critical need to develop methods for selecting optimal source domain datasets and to achieve truly automated tagging, to address these current limitations and progress toward cost-free automated tag generation.

In July 2023, Plant phenomics published a research article titled “EasyDAM_V3: Automatic fruit labeling based on optimal source domain selection and data synthesis via a knowledge graph.”

In an effort to promote automatic fruit labeling with High efficiency and no cost, this study introduces EasyDAM_V3, a new approach that combines optimal source domain selection with synthetic dataset generation. EasyDAM_V3 aims to address two primary challenges: selecting the most appropriate source domain fruit datasets for image translation and reducing the manual annotation cost in the target domain.

The first aspect of EasyDAM_V3 involves a systematic selection of source and target domain datasets for image translation models. This process uses a multidimensional spatial feature model, which enables the identification of the most suitable source domain dataset which can match multiple target domain fruits. The selection is based on the analysis of phenotypic characteristics such as shape, color and texture across various fruit datasets.

For example, in the study pears were identified as the optimal source domain for translating images to target domains such as citrus, apples and tomatoes. This determination was made by a clustering algorithm and multidimensional feature space analysis, which ensures a higher fidelity in translation generalization. The second aspect focuses on building a knowledge graph to generate synthetic datasets with accurate label information.

EasyDAM_V3 uses transparent background fruit image translation and an anchor-free detector for pseudo-label self-learning. This innovative approach can handle fruits of different scales and shapes, improving the final label generation accuracy.

The experimental validation of EasyDAM_V3 involved citrus, apple and tomato as the target domains. The process consisted of three main parts: using multidimensional feature quantization and spatial reconstruction to select the optimal source domain fruits, inputting these source fruits into the CycleGAN model for target domain image generation, and using these images to construct synthetic datasets.

These datasets were then used to train an anchor-free detector-based fruit detection model. Results of these experiments showed that EasyDAM_V3 could successfully translate and generate labels for the target domains using pears as the source domain, with high average precision rates of around 90%. This demonstrates EasyDAM_V3’s effectiveness in addressing both challenges of optimal source domain selection and reducing manual annotation costs.

In summary, the approach outlined by EasyDAM_V3 not only improves the applicability and domain adaptability of automatic labeling algorithms, but also represents an important step towards achieving efficient, cost-effective solutions in agricultural AI and smart orchard management.