Feature enhancement guided network for yield estimation of high-density jujube

Area of Science:

• Agricultural Science
• Computer Vision
• Artificial Intelligence

Background:

• Accurate jujube yield prediction is crucial for orchard management and resource allocation.
• Traditional object detection methods struggle with dense jujube populations due to occlusion and labor-intensive labeling.
• Existing techniques are not robust in real-world conditions with varying scales, backgrounds, and illumination.

Purpose of the Study:

• To develop a robust and efficient image-based method for predicting jujube yield in high-density orchards.
• To overcome the limitations of traditional methods in handling occlusion and dense target settings.
• To advance the application of AI in high-density agricultural target recognition.

Main Methods:

• Developed a feature enhancement guided network for jujube counting and yield estimation.
• Proposed a novel label representation method based on uniform distribution for improved object characterization.
• Integrated a feature enhancement module to guide a density regression module for accurate predictions.

Main Results:

• The proposed method achieved high accuracy in estimating jujube numbers with a Mean Absolute Error (MAE) of 9.62 and Mean Squared Error (MSE) of 22.47.
• Experiments were conducted on a dataset of 692 images with 40,344 jujubes.
• The method significantly outperformed state-of-the-art techniques in jujube yield estimation.

Conclusions:

• The developed image-based technique offers an efficient solution for predicting jujube yield.
• This research contributes to advancing AI applications in agriculture and forestry for high-density target recognition.
• The technique aims to enhance agricultural automation and optimize resource allocation.