MangoNet: A deep semantic segmentation architecture for a method to detect and count mangoes in an open orchard

This work presents a method for detection and counting of mangoes in RGB images for further yield estimation. The RGB images are acquired in open field conditions from a mango orchard in the pre-harvest stage. The proposed method uses MangoNet, a deep convolutional neural network based architecture for mango detection using semantic segmentation. Further, mango objects are detected in the semantic segmented output using contour based connected object detection. The MangoNet is trained using 11,096 image patches of size 200×200 obtained from 40 images. Testing was carried out on 1500 image patches generated from 4 test images. The results are analyzed for performance of segmentation and detection of mangoes. Results are analyzed using the precision, recall, F1 parameters derived from contingency matrix. Results demonstrate the robustness of detection for a multitude of factors such as scale, occlusion, distance and illumination conditions, characteristic to open field conditions. The performance of the MangoNet is compared with FCN variant architectures trained on the same data. MangoNet outperforms its variant architectures.