Implementation of Convolutional Neural Network in Detecting Avocado Ripeness Level

Squeezing avocados to determine ripeness can cause physical damage or bruising, reducing the fruit’s quality and resulting in losses for sellers and buyers. This research aims to develop an Android-based mobile application to detect avocado ripeness based on skin color, avoiding physical damage to the fruit. The study uses three simple Convolutional Neural Network architectures to evaluate the algorithm’s ability to detect avocado ripeness. The dataset includes 385 images across four classes: immature, half-ripe, ripe, and overripe (74 images each), and an additional 89 images for the non-avocado class. The model was trained with learning rates of 0.001, 0.0001, and 0.00001. The architecture with the most convolutional layers achieved the best performance with a 0.001 learning rate, yielding a test accuracy of 94.15%, a test loss of 19.28%, and an F1-score of 94.0%. The best model was then converted to TFLite format and successfully integrated into an Android application that functions effectively.