Abstract:
The apple is among the most widely consumed fruits globally due
to its high nutritional value and longer shelf life. Consequently,
there is a need to monitor post-harvest health, which has important
applications in grading, packaging and transportation. These tasks
require the accurate computation and prediction of fruit size; however, there is a lack of publicly available datasets that the researcher
can use to generalize it over apples. To bridge this gap, we introduce the AppleV dataset, which consists of images of 200 apples
taken from different angles and sums up to 2000 images with their
corresponding volume measurement. This work also explores three
traditional approaches(Analytical method, Support Vector Machine
Regression(SVMR), and Dynamic weighting) and five deep learning approaches(Custom model, EfficientNetV2S, VGG16, Xception,
and ResNet50) for volume estimation. Traditional approaches use
handcrafted features to estimate volume, whereas deep learning
approaches learn these features automatically. To mitigate the risk
of overfitting caused by the large number of trainable parameters in
state-of-the-art deep learning models, pre-trained models trained
on the ImageNet dataset are utilized. The results indicate that while
the custom model demonstrated superior performance on the training data, ResNet50 outperformed all other methods on the AppleV
dataset in terms of validation data accuracy. This work tries to offer
a standard dataset specifically for the volume estimation task and
aims to advance agricultural technology by automating volume
estimation for apples based on image data, facilitating improved
quality assessment and has various applications in industries such
as grading and sorting.
