Applying artificial intelligence in durian fertile lobe detection: Attention-Residual Unet and Test Time Augmentation algorithm

The key factor in durian fruit trading is ripeness. Several studies have been conducted on non-destructive durian maturity classification using near-infrared (NIR) spectroscopy. However, most of these studies manually determined the most accurate measurement position, which was the durian's main fertile lobe center. This research aims to automate the stage of detecting this position of the durian by using UNet segmentation method, which leverages differences in rind texture between the center of the main fertile lobe and other areas (lobe grooves and stems), prior to conducting NIR measurements. The rough and non-uniform surface of the durian rind presents a significant challenge for segmentation. However, the large size of the durian spines in the main fertile lobe serves as an identification characteristic for the segmentation model. This study uses the Ri-6 durian in Vietnam as the samples for the experiment. The model was developed using three architectures: Unet, Attention-Unet and Attention-Residual Unet. According to the analysis results on test set, Unet, Attention-Unet and Attention-Residual Unet algorithms achieved %accuracy of 78.22%, 81.34%, 82.89% and %intersection over union of 79.49%, 80.47%, 80.72%, respectively. After that, the model was further enhanced using the test time augmentation algorithm, improving the %accuracy to 85.24%, 85.68%, 86.85% and %IoU to 81.65%, 82.03% and 83.12%. Among the three architectures, the Attention-Residual-Unet model demonstrated the highest efficiency in detecting the center of the durian’s main fertile lobe for non-destructive durian maturity classification. This method can be applied to the development of an automatic durian’s maturity classification machine, which would save time and improve economic efficiency.