Boosting Pineapple Maturity Classification: Impact of Data Augmentation and Visual Transformer Integration with Transfer Learning

Pineapple is a tropical fruit with varying degree of ripeness. It plays an important role in tribal agricultural markets. The effective pineapple maturity grading is essential for optimizing distribution and maximizing profits of tribal farmers. This study utilizes a pineapple dataset from Kaggle which include raw and ripe pineapple images. These raw and ripe are categorized into four maturity classes - fully ripe, medium ripe, partially ripe, and unripe - using Hue, Saturation, and Value (HSV) color space technique. Furthermore, the original dataset further divided into four specialized datasets: cluttered-unbalanced (d₁), cluttered-balanced (d₂), fine-tuned-unbalanced (d₃), and fine-tuned-balanced (d4). In this study, we explore the effectiveness of four Transfer Learning Models (TLMs): DenseNet121, VGG19, MobileNetV2, and InceptionV3, in conjunction with Visual Transformer (VT) technology for pineapple maturity classification. In addition, this work introduced four Hybrid models popularly named as VT-TLMs (VT-DenseNet121, VT-VGG19, VT-MobileNetV2 and VT-InceptionV3). These Hybrid models make the best use of the features that are extracted from the final layers of TLMs to capture complex features, detailed patterns and long-range dependencies inherent in image classification tasks. The proposed TLMs can decide maturity of pineapples with an accuracy ranging from 91.49% to 93.64%. Moreover, VT integration with TLMs obtained better accuracy ranging between 94.43% and 99.57%. Additionally, VT-DenseNet121 performs exceptionally well, with an average accuracy of 99.57%, when compared to previous research in the field.