Evaluation and Comparative Analysis of Feature Extraction Methods on Image Data to increase the Accuracy of Classification Algorithms

Abstract

Manual selection of fresh fruit has been identified as a significant challenge for the agricultural
sector due to its time-consuming nature and potential for inconsistent categorization. This process
requires human labour to visually inspect and sort fruits, leading to variability and inefficiencies
in the sorting process. This research proposes a low-cost alternative using intelligent fruit selection
systems based on computer vision techniques to address these issues. These systems aim to
automate the process of fruit selection, improving efficiency and consistency in categorizing fruits
such as apples, bananas, and oranges. A critical step in developing such intelligent systems is the
feature extraction process. Feature extraction is essential in classification, especially for data
sources in the form of images. It involves identifying and isolating relevant information from the
images that classification algorithms can use to distinguish between different fruit categories. If
the feature extraction process fails to capture the correct information, the performance or accuracy
of the classification algorithm will be negatively impacted. This research compares three different
methods for extracting features from fruit images to determine which method yields the highest
accuracy for fruit classification. The feature extraction methods evaluated were Grayscale Pixel
Values, Mean Pixel Value of Channels, and Extracting Edge Features. The classification algorithm
used in this research is the Convolutional Neural Network (CNN) algorithm. CNNs are well-suited
for image classification tasks due to their ability to learn hierarchical feature representations from
the input images automatically. By comparing the performance of the CNN classifier using the
three different feature extraction methods, this research aims to identify the method that provides
the highest level of accuracy in classifying fruit images. By conducting this comparative analysis,
the research provides insights into the most effective feature extraction techniques for improving
the performance of computer vision-based fruit selection systems, ultimately contributing to more
efficient and consistent fruit categorization in the agricultural sector. The result revealed that the
Grayscale achieved the highest validation accuracy (99.94%) and the lowest validation loss
(0.44%).