DEVELOPMENT OF AN AUTOMATED DEBRIS DETECTION SYSTEM FOR WILD BLUEBERRY HARVESTERS USING A CONVOLUTIONAL NEURAL NETWORK TO IMPROVE FRUIT QUALITY
Das, Anup Kumar
MetadataShow full item record
Improving wild blueberry fruit quality has become increasingly important to producers due to the tightening profit margin facing the industry. The continuous development of field management practices (i.e., application of fungicides, herbicides, fertilizers, pollination, and pruning, etc.), has improved the wild blueberry field vegetation by significant increases in plant densities, plant height, and fruit yield. This increased plant debris causes additional leaves and stems to enter into the fruit storage bins during mechanical harvesting resulting in a potential reduction of fruit quality. An experimental image dataset (1000 images) was collected from a mechanical harvester in two commercially managed fields in central Nova Scotia. Three different deep learning algorithms (YOLOv3, YOLOv3-Tiny, and YOLOv3-SPP) were implemented and compared for developing the real-time debris detection system. The image dataset was augmented using five different color-based data augmentation techniques (sharpening, brightness, contrast, gamma correction, saturation). T1 dataset containing 2000 images was prepared by changing sharpness, brightness, contrast, gamma, and saturation value of 1000 images by factors of 10, 20, 20, 1.3 and 1.5 respectively and mixing with 1000 experimental images. Similarly, T2 dataset containing 3000 images was prepared by changing the gamma value of images by factors of 0.70, 0.80, 0.90, 1.10 and 1.20 and mixing with experimental images and T1 dataset. YOLOv3-SPP achieved 73.03% of mAP and 74.38% of mAP after training on T1 and T2 dataset. YOLOv3-SPP improved mAP by 5.89% from previous accuracy (YOLOv3-SPP: mAP = 68.49%). The two best performing models (Model-1: YOLOv3-SPP, mAP: 74.38%; Model-2: YOLOv3-SPP, mAP: 73.03%) were tested on four different types of CPU, GPU, and embedded based computers. A 4x2 factorial design was used to choose the most appropriate hardware and model for developing the system. Results determined YOLOv3-SPP (mAP: 73.03%) operated using Intel® Core™ i9-7900X CPU @ 3.30 GHz, GeForce RTX™ 2080 Ti @ 1665 MHz on a desktop computer achieved the fastest detection (33.30 ms) with the highest average frame rate (30.03 FPS). Model-1 and Model-2 achieved 85.90% and 86.10% of validation mAP respectively at 0.10 IoU and 0.10 confidence threshold. A paired t-test was implemented to investigate the difference between automatic detection with ground truths. Results showed that the developed system successfully detected green berries, ripe berries and leaves during lab evaluation at 5% level of significance. However, stems, and dirt classes were not successfully detected from ground truths at 5% level of significance. In this study YOLOv3-SPP was implemented for developing debris detection system which performed comparatively better than YOLOv3 or YOLOv3-Tiny. This system can be incorporated in a control system to automate brush adjustment on the basis of feedback from conveyors of mechanical wild blueberry harvesters. This system can be a valuable addition for enhancing berry cleaning efficiency and improving fruit quality.