Abstract
Deep learning-based measurement techniques are increasingly supporting postharvest supply chains by automating tasks such as fruit inspection, grading, and maturity assessment. Developing lightweight yet accurate models that operate reliably across diverse visual conditions remains challenging. This study addresses three objectives: (i) to design a deep learning system for multi-class sapota quality grading, (ii) to incorporate attention mechanisms to improve feature focus with minimal parameter increase, and (iii) to achieve high accuracy and robustness under varied background conditions. An attention-enhanced MobileNetV2–Convolutional Block Attention Module (CBAM) model is proposed to meet these objectives. The model integrates channel and spatial attention modules to refine feature extraction while maintaining compactness suitable for mobile devices. A digitized dataset, comprising 6,010 training images and 4,806 cross-environment test images (blue, white, grass and soil backgrounds), is used to evaluate measurement consistency. Experimental results show that the proposed system achieves 94.92 % accuracy with only 3.03 M trainable parameters, outperforming the baseline MobileNetV2 by +4.25 % accuracy. Cross-environment experiments yield 74 % accuracy, demonstrating robustness against background-induced measurement variability. An inference time of 25 ms on a Graphics Processing Unit (GPU) indicates potential for real-time applications. Measurement interpretability is validated using Gradient-weighted Class Activation Mapping (Grad-CAM), confirming consistent focus on size, texture, and defect regions. Overall, this study demonstrates that integrating CBAM with lightweight architectures significantly enhances measurement precision and generalization, contributing to scalable automated fruit quality inspection systems.