ASE-UNet: An Orange Fruit Segmentation Model in an Agricultural Environment Based on Deep Learning

Abstract

Fruit picking robot requires a powerful vision system that can accurately identify the fruit on the tree. Accurate segmentation of orange fruit in orchards is challenging because of the complex environments due to the overlapping of fruits and occlusions from foliage. In this work, we proposed an image segmentation model called ASE-UNet based on the U-Net architecture, which can achieve accurate segmentation of oranges in complex environments. Firstly, the backbone network structure is improved to reduce the down-sampling rate of orange fruit images, thereby retaining more spatial detail information. Secondly, we introduced the Shape Feature Extraction Module (SFEM), which at enhancing the ability of the model to distinguish between the fruits and backgrounds, such as branches and leaves, by extracting shape and outline information from the orange fruit target. Finally, an attention mechanism was utilized to suppress background channel feature interference in the skip connection and improve the fusion of high-layer and low-layer features. We evaluate the proposed model on the orange fruit images dataset collected in the agricultural environment. The results showed that ASE-UNet achieves IoU, Precision, Recall, and F1-scores of 90.03, 96.10, 93.45, and 94.75%, respectively, which outperform other semantic segmentation methods, such as U-Net, PSPNet, and DeepLabv3+. The proposed method effectively solves the problem of low accuracy fruit segmentation models in the agricultural environment and provides technical support for fruit picking robots.