Cocoa pods ripeness estimation, using convolutional neural networks in an embedded system

Abstract

A correct cocoa harvest involves determining a pod maturity. However, this farm activity is usually handmade, using criteria such as Size and Color of the pod; those characteristics differ according to the cocoa variety, making it difficult to standardize. For this reason, this work proposes an automated method to simplify the number of variables to develop a portable, low-cost, and custom-made tool, which makes use of a convolutional neural network to indicate whether a cocoa pod is found it at the right time to harvest. The main results of this work are: 1) the construction of three labeled data sets (1992 images each), and 2) we developed an embedded system with a 34.83% mAP (mean Average Precision) accuracy. Finally, variance analysis demonstrates that image size (i.e., 4033x4033 p, 1009x1009 p, and 505x505 p) does not affect accuracy.