A unified mass estimation pipeline from a single 2D image for fruits with diverse shapes using sparsely selected saliency-driven deep feature embeddings

Accurate and non-destructive mass estimation is essential in modern agricultural automation, yet existing approaches still face challenges in providing consistent and lightweight solutions for fruits with diverse geometries. Traditional geometric approximations and handcrafted visual features may be limited when dealing with complex structural variations in real-world production. This study presents a calibrated saliency-based deep feature framework for estimating fruit mass from a single 2D top-view image. Multi-scale hierarchical saliency maps are used to generate morphology-related embeddings for regular and irregular fruit shapes. The fused saliency representation is encoded into a 1024-dimensional deep feature vector, from which informative dimensions are selected via sparse LASSO, and the final mass prediction is obtained using optimized Ridge Regression. The proposed method is evaluated on avocado, persimmon, and starfruit datasets under fruit-specific 5-fold cross-validation, achieving a maximum R 2 of 98.47% and a minimum MAPE of 2.43% across the evaluated datasets. The results indicate that the same saliency–sparsity–regression pipeline can be effectively calibrated across fruit datasets with markedly different morphologies while maintaining a compact feature representation. Overall, the framework provides a lightweight 2D imaging solution with no task-specific deep fine-tuning for empirical non-destructive fruit mass estimation in precision agriculture.