Deformable Convolutional Networks-Based Instance-Specific Coefficients of On-tree Mango Segmentation

This research does provide an approach for the detection of mangoes on trees with the use of deformable convolutional networks concentrating on imaging at nighttime. The method used in the mango detection and segmentation is YOLACT++ which is critical in estimating the sizes of fruits in remote sensing applications. The evidence from experiments shows the effectiveness of the proposed model in various lighting conditions. Despite this, it means that this model yields high detection accuracy particularly at night where confidence scores are above 0.90 for light and noise. During the daytime, images capture by the model show better segmentation accuracy and detection probabilities of the model. The average precision of detection (AP) for bounding boxes is in the 53.45 – 79.67 range, with AP exceeding 80% at lower Intersection over Union (IoU) values (0.5–0.7). However, the localization of the mangoes has remained a task when correctly estimating them using higher IoU thresholds of 0.75 to 0.95. These achievements further confirm that YOLACT++ can perform mango instance segmentation and suitable for real-life uses, including agricultural surveillance and fruit measurement.