Intelligent Layout Segmentation and Classification of Packaging Labels Using Enhanced DeepLabv3+ with Multi-Scale Feature Optimization

Automated analysis of packaging labels poses significant challenges in computational efficiency and adaptability to complex layouts. This study proposes a lightweight segmentation framework integrating depthwise separable convolution and multi-scale feature enhancement. The conventional Xception backbone is replaced with MobileNetv3-small, reducing model parameters by 68% (from 13.1M to 4.2M). An enhanced Atrous Spatial Pyramid Pooling (ASPP+) module with multi-level dilation rates (2, 4, 8, 12, 16) effectively addresses multi-scale feature extraction for sub-millimeter elements (e.g., 0.2mm pharmaceutical batch codes), high-density text, and intricate graphics. A specialized dataset of 1,000 high-resolution packaging labels (average 3500×5200 pixels) across food, pharmaceutical, and cosmetic categories is constructed, featuring pixel-level semantic annotations and eight industrial layout classes, partitioned into training, validation, and test sets at a 7:2:1 ratio. The proposed model achieves 81.3% mIoU on the test set, outperforming FCN (71.5%), PSPNet (75.0%), and the original DeepLabv3+ (77.6%) by 9.8, 6.3, and 3.7 percentage points, respectively. Inference speed reaches 0.32 seconds per frame (NVIDIA RTX 4080 Ti). For critical industrial metrics, the framework attains 97.5% accuracy in overlapping element segmentation, constrains edge errors for 0.2mm batch codes to < 1 pixel (ISO 15378 compliant), and reduces false detection rates for semi-transparent labels by 18.3% (F-score: 0.891). Integrated with a GoogLeNet-based layout classifier (90.56% accuracy), this system provides an efficient, high-precision solution for automated compliance verification and intelligent packaging design.