JoWUA

Abstract

Precise identification of fruits is an important requirement for intelligent agriculture automation for efficient harvesting and grading. The performance of computer vision algorithms may be hampered by various factors such as insufficient training samples, class imbalance, presence of noises, illumination conditions, and occlusions. To mitigate the problems, this work suggests an approach that uses the combination of GANs data augmentation and effective image denoising for the purpose of increasing the quality of the fruit dataset. Four GAN models, which include InfoGAN, CycleGAN, StyleGAN, and Conditional GAN (cGAN) have been compared for the task of synthetic images creation, while four image denoising models have been studied, including Pixel Density-based Trimmed Median Filter (PDTMF), Color Wiener Filter (CWF), Guided Box Filter (GBF), and Structural Interval Gradient Filter (SIGF). The experimental evaluation was carried out on a dataset with more than 2,000 images of 18 classes of fruits. The assessment was done with the help of Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM). The cGAN proved to have the best performance among all tested augmentation models with PSNR = 11.07 dB and SSIM = 0.815. In the context of image denoising, PDTMF surpassed the other filters in performance, yielding a PSNR value of 31.26 dB and SSIM of 0.917. The combination of GAN-based data augmentation with adaptive denoising was instrumental in boosting the quality of the images as well as making the datasets more robust for future use in object detection under adverse environmental conditions. The proposed framework offers a scalable preprocessing solution for precision agriculture and establishes a strong foundation for future implementation of real-time fruit detection using deep learning architectures such as YOLOv8.