Investigating the Effects of Pre-Trained Deep Learning Models and Fusion Techniques on Fruit Segmentation Performance

In recent years, interest in the automation of agricultural and industrial processes has increased. Fruit segmentation is essential for quality control, harvest optimization, and yield estimation. However, some challenges arise in automatic fruit segmentation systems, such as light conditions, image quality, environmental factors, and variations in fruit appearance. To over-come these challenges, we proposed an approach by investigating the effects of pretrained deep learning models with late fusion techniques. Firstly, we designed a late fusion module focusing on the information fusion in the last two layers of the deep learning models. Then, we integrated this late fusion module into the two parallel deep learning models that process the input images independently. Finally, the fruit segmentation output is produced based on the combined information in the late fusion module. The proposed method achieved an IoU score of 79.13% on the Minneapple dataset. Compared to CNN and Transformer models, the proposed method achieved a performance improvement of 9.42% and 3.29%, respectively.