Ju-LiteMobileAtt: A lightweight attention network for efficient jujube defect classification

Abstract

Surface defect detection of organic jujubes is critical for quality assessment. However, conventional machine vision lacks adaptability to polymorphic defects, while deep learning methods face a trade-off—deep architectures are computationally intensive and unsuitable for edge deployment, whereas lightweight models struggle to represent subtle defects. To address this, we propose Ju-LiteMobileAtt, a high-precision lightweight network based on MobileNetV2, featuring two key innovations: First, the Efficient Residual Coordinate Attention Module (EfficientRCAM) integrates spatial encoding and channel interaction for multi-scale feature capture; Second, the Cascaded Residual Coordinate Attention Module (CascadedRCAM) refines features while preserving efficiency. Experiments on the Jujube12000 dataset show Ju-LiteMobileAtt improves accuracy by 1.72% over baseline while significantly reducing parameters, enabling effective real-time edge-based jujube defect detection.