A Semantic Segmentation Method for Segmenting Chicken Parts Based on a Lightweight DeepLabv3+

Abstract

Research on poultry part partitioning techniques is crucial for the advancement of automated poultry partitioning equipment. In this study, a semantic segmentation method for chicken parts, based on a lightweight DeepLabv3+, was introduced to cater to real-time and precise requirements of segmenting varying poultry sizes. Initially, the backbone network was replaced with an improved lightweight MobileNetV2, enhancing the predictive speed and decreasing computational parameters. Subsequently, the SENet was incorporated, enhancing the capacity to discern high-level features and negate irrelevant information. Furthermore, two shallow feature layers of different scales were integrated into the decoder, augmenting the richness of shallow features and mitigating inaccuracies at segmentation edges. Finally, the Dice Loss and Cross Entropy Loss (CE Loss) functions were combined to minimize the imbalance between positive and negative samples. Experimental findings demonstrated that the lightweight DeepLabv3+ improved the MIoU (Mean Intersection over Union) and MPA (Mean Pixel Accuracy) scores of the original model by 5.42% and 3%, respectively, and amplified the detection speed by 1.89 times. Remarkably, the model size was a mere 10.95% of the original, indicating substantial enhancements in segmentation accuracy and detection speed. Therefore, the proposed algorithm could potentially provide certain technical insights for automatic segmentation of different poultry.