MASPP and MWASP: multi-head self-attention based modules for UNet network in melon spot segmentation

Sweet melon, and in particular, spotted melon, is one of the most profitable fruit crops for farmers in the international market. As the spot ratio impacts the melon's visual appeal, it plays a significant role in shaping consumers' initial impressions and influencing their decision to purchase a spotted melon. However, accurately determining the spot area on a melon's skin is challenging due to the diverse sizes and colors of these spots among different types of melons. In this study, the novel networks based on UNet model have been proposed to accurately determine the spot area on melon skins after harvesting. First, Mask R-CNN model was employed to isolate the melons from unwanted objects and backgrounds. Then, the novel variants of the Atrous Spatial Pyramid Pooling (ASPP) and Waterfall Atrous Spatial Pooling (WASP) were developed based on the multi-head self-attention (MHSA) approach to efficiently enhance the original structures. Finally, the proposed modules were integrated into VGG16-UNet network to segment melons' spots on its skin. The experimental results demonstrate that the proposed methods yielded promising outcomes, achieving a mean IoU of 89.86% and an accuracy of 99.45% across all classes. Moreover, it outperformed other existing models.