Detection method for munage grape clusters and abnormal berries under color-similar backgrounds

Accurate detection of Munage grape clusters and abnormal berries in field scenes is hindered by several challenges: mature berries often exhibit colors similar to those of branches and leaves; cluster-level large targets coexist with medium- and small-scale abnormal berry targets; local abnormal cues within full-berry bounding boxes are easily diluted by responses from normal berry skin and waxy bloom; and shallow-level textures, specular highlights, and adjacent berry boundaries may induce false detections. To address these challenges, this study proposes YEIS, a YOLO11n-based detection method for Munage grape clusters and abnormal berries under color-similar backgrounds. Built upon YOLO11n, the proposed method first introduces EMBSFPN to construct multi-scale candidate features, thereby alleviating the scale-representation discrepancy between cluster-level targets and berry-level abnormal targets. Second, an intra-berry frequency-local evidence decoupling module, IB-FLED, is designed to enhance local abnormal cues within full-berry detection boxes through low-frequency appearance estimation, local residual modeling, and morphology-aware response branches. Finally, a semantic-guided recall compensation module, SGRCM, is developed to constrain P3 detail compensation using P4 semantic information refined by IB-FLED, reducing the interference of shallow-level textures, waxy bloom, and specular highlights in abnormal berry localization. Three random-seed experiments were conducted on a self-built field dataset. The results show that YEIS achieves Precision, Recall, mAP₅₀, mAP₇₅, and mAP_50-95 values of 83.21%, 79.97%, 88.34%, 83.27%, and 76.91%, respectively. Compared with YOLO11n, the overall mAP_50-95 is improved by 1.71 percentage points, and the mAP_50-95 for lesion-like abnormal berries is increased by 1.61 percentage points. For scar-like abnormal berries, the F1-score and mAP_50-95 are improved by 3.01 and 3.41 percentage points, respectively. Meanwhile, the number of model parameters is reduced from 2.583 M to 2.149 M, corresponding to a reduction of 16.8%. The proposed method improves the detection and localization of abnormal berries under color-similar backgrounds while reducing the parameter count relative to YOLO11n, providing a front-end visual detection approach for digital monitoring, grape-cluster localization, and abnormal-berry recognition in Munage vineyards.