DefinePK

Abstract

Detection of fruit ripeness properly is important in improving the quality of the fruits after harvesting and increasing the time of harvesting the fruit as well as minimizing food waste that is utilized in agriculture practices nowadays. The existing methods doing things by hand, through the comparison of colors and chemical analysis could be highly subjective, time consuming, and impractical on large scale. To avoid these disadvantages, this study proposes a new hybrid deep learning case that encourages to integrate YOLOv6 (You Only Look Once version 6) with Attention Transformer into mutually training to identify the fruits location and that of ripeness’s of oranges in an environment within a natural orchard setup. The sample of the annotated pictures of orange that had been taken (N=5000 images) in Sargodha, Pakistan, of four stages of ripeness (unripe and partially ripe, ripe or overripe), under various conditions (different lighting and background aspects). The YOLOv6 backbone makes it possible to recognize the objects during the live video and the attention mechanism, which is integrated into it, adds to the competencies of the active to underline finer details of color, brightness, and surface texture that are necessary to outline the boundary between the related stages of maturity. The layer of attention rectifier corrects the region specific features that have been determined by YOLOv6 adding to intra class distinction to itself and even capacity to counter the environmental noises, confusion, and even shadow. The experimental results that show that the proposed model is significantly superior to the baseline networks, such as ResNet-50 and individual YOLOv6, is because the mean Average Precision (mAP) of 97.4 %, Precision of 98.1 %, Recall of 96.8 %, and F1-score of 97.4% measurements are high values. Classify by ripeness recovery has the same performance at all the ripeness stages with minimum false recovery between partially and ripe categories- the sensitivity of the fineness of the models. The hybrid frame in addition consists of real-time (approximately 59 FPS) inference which is proof that it can be employed in automated tests of fruit-sorting as well as high precision farming. According to the findings, when the attention based buildings are collected together with the latest detection frameworks, it becomes clear that the accuracy and interpretation of the methods of determining the maturity of fruits are enhanced. Moreover, further research will be invested in multi-spectral and hyperspectral photography, time tracking, and another stage advancement of the model in the lived environment consisting of dense orchard and be able to use it in more other branches.