How Can a Transformer Spot Camouflaged Fruits on Trees?

Identifying fruits on trees is key for disease prevention, yield estimation, and automated harvesting. The challenge? In orchards fruits often blend with leaves, branches, and light, making detection hard. To tackle this, researchers developed FCNet, a transformer‑based system that “sees” the whole tree context. The model uses PVTv2 for multi‑scale feature extraction, then enhances them with an ASPP module that captures details at various resolutions and attention mechanisms that highlight important regions. Finally, a special Position‑based Layer Normalization refines fruit edges for sharper segmentation. Tested on mango images, FCNet outperformed all rivals, boosting metrics like S‑measure, IoU, and F‑measure by several percentage points. Ablation studies confirmed that both feature enhancement and final normalization are crucial for performance. In practice, this could make harvesting robots more reliable and help farmers monitor orchards with fewer mistakes.

FCNet builds on a Vision Transformer variant called PVTv2, which introduces a hierarchy of tokens at decreasing resolutions, allowing capture of both fine details and global context. ASPP (Atrous Spatial Pyramid Pooling) employs dilated convolutions with multiple dilation rates, creating multi‑scale receptive fields without increasing parameter count, ideal for objects of varying size like fruits. The attention module combines channel‑wise and spatial‑wise attention, enhancing discrimination between fruit and background. Position‑based Layer Normalization (PLN) normalizes activations while considering pixel spatial positions, reducing bias from illumination or perspective changes. Training used a mango image dataset with pixel‑wise annotations, employing a combined BCE and Dice loss to balance precision and recall. Results show an average 4‑6% boost on segmentation metrics compared to UNet, DeepLabV3+, and SegFormer, confirming the effectiveness of the multi‑scale approach and positional refinement. Ablation studies revealed that removing ASPP drops S‑measure by 2.43 points, while omitting PLN penalizes IoU by 2.73 points, highlighting both components' importance for robustness in visually camouflaged scenarios.