DALF-YOLO: Cherry Tomato Ripeness Detection in Complex Environments via Frequency-Domain Dynamic Attention and Local–Global Multiscale Fusion

Rising labor costs have accelerated the adoption of harvesting robots, for which accurate fruit recognition and ripeness detection are essential. However, cherry tomato ripeness detection in unstructured field environments remains challenging owing to complex illumination, leaf occlusion, fruit overlap, background similarity, and color variation. To address these challenges, we propose DALF-YOLO, an improved YOLO26n-based method that integrates frequency-domain dynamic attention and local–global multi-scale fusion for cherry tomato recognition and ripeness detection. Specifically, an AMFF module is designed to enhance interactions between high-level semantic features and low-level details through cross-scale skip connections, bidirectional weighted fusion, and SimAM attention, thereby improving multi-scale detection. A lightweight LGAM module is embedded in the backbone, employing Ghost Convolution and decoupled fully connected attention to reduce computational cost while strengthening global feature representation and robustness to complex backgrounds. In addition, an FSDSA module is developed to separate texture and contour information in the frequency domain and perform dynamic sampling and fitting in the spatial domain, improving the detection of irregular fruit edges and overlapping tomatoes. These designs enhance feature extraction while maintaining a lightweight architecture. Experimental results show that DALF-YOLO achieves an average accuracy of 94.3%, outperforming the baseline by 4.5 percentage points, with an average inference time of 7.58 ms. FP16-based edge deployment further achieves 20-27 FPS, demonstrating the feasibility of real-time cherry tomato ripeness detection for selective robotic harvesting.