Articles
Machine vision with deep learning for in-orchard mango fruit sizing and size distribution
Article number
1395_42
Pages
317 – 324
Language
English
Abstract
Tree fruit harvest load is a function of fruit number and fruit size.
Forecast of fruit load facilitates management of harvesting, packhouse function and marketing.
The traditional method of fruit size assessment involves manual assessment, a labour-intensive process.
On-tree fruit size estimation using machine vision has been enabled by convolution neural network (CNN) based object detection and segmentation methods and by the advent of depth cameras.
Several issues limiting technology application to this application were considered: i) hardware: another depth camera was benchmarked to eight previously characterised cameras; ii) measurement accuracy across the field of view was considered; iii) the accuracy of a filter for fruit occlusion was assessed; iv) two sizing approaches were trialled; and v) fruit sizing was undertaken using images captured from a vehicle moving between rows at 7 km h‑1. Sizing of lineal dimensions was achieved to an RMSE of <4 mm for fruit length for detections near the centre of imagery captured from a vehicle moving at 7 km h‑1 with using a neural network-based fruit segmentation method and exclusion of partly occluded fruit based on fitted ellipse attributes.
Forecast of fruit load facilitates management of harvesting, packhouse function and marketing.
The traditional method of fruit size assessment involves manual assessment, a labour-intensive process.
On-tree fruit size estimation using machine vision has been enabled by convolution neural network (CNN) based object detection and segmentation methods and by the advent of depth cameras.
Several issues limiting technology application to this application were considered: i) hardware: another depth camera was benchmarked to eight previously characterised cameras; ii) measurement accuracy across the field of view was considered; iii) the accuracy of a filter for fruit occlusion was assessed; iv) two sizing approaches were trialled; and v) fruit sizing was undertaken using images captured from a vehicle moving between rows at 7 km h‑1. Sizing of lineal dimensions was achieved to an RMSE of <4 mm for fruit length for detections near the centre of imagery captured from a vehicle moving at 7 km h‑1 with using a neural network-based fruit segmentation method and exclusion of partly occluded fruit based on fitted ellipse attributes.
Authors
C. Neupane, K.B. Walsh, A. Koirala
Keywords
fruit sizing, fruit segmentation, machine vision, mango, occlusion filtering, perspective, shape fitting
Groups involved
- Division Temperate Tree Nuts
- Division Vine and Berry Fruits
- Division Temperate Tree Fruits
- Division Precision Horticulture and Engineering
- Division Plant-Environment Interactions in Field Systems
- Division Tropical and Subtropical Fruit and Nuts
- Working Group Precision Management of Orchards and Vineyards
Online Articles (59)