Improved hybrid feature extractor in lightweight convolutional neural network for postharvesting technology: automated oil palm fruit grading.

Grading of oil palm fresh fruit bunches (FFB) plays a vital role in the postharvest operation as it directly influences the extraction rate of oil palm, thereby ensuring quality control in the estate and mill. Currently, manual grading based on visual assessment is employed, but it has limitations mainly due to subjective judgment-influenced factors such as visual resemblance, light intensities, and differences in colors across ripeness categories. Hence, an automated oil palm fruit grading system in postharvest technology is proposed to enhance the grading process and improve productivity while maintaining operational cost efficiency. This work involves developing an improved object detection model based on the You Only Look Once model to accurately identify four grades of oil palm FFB, namely, ripe, unripe, underripe, and overripe. The proposed model incorporated several improvements, including a hybrid feature extractor comprising mobile inverted bottleneck module and densely connected neural network. Additionally, it employs a spatial pyramid pooling structure to expand the receptive field and utilizes the complete intersection over union function for bounding box regression. The results indicate that the proposed model obtains a remarkable mAP of 94.37% and an F1-score of 0.89. Besides, the model performs real-time detection at a faster rate of 4.8 FPS on a limited-capacity embedded device, NVIDIA Jetson Nano. The comprehensive experimental results confirm the superiority of the proposed model over various detection models. [ABSTRACT FROM AUTHOR]