Recognition of Occluded Goods under Prior Inference Based on Generative Adversarial Network

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Intelligent retail systems face challenges in accurate product recognition due to occluded goods and high visual similarity between items.
• Existing methods struggle with feature loss from occlusions and distinguishing between visually similar products, impacting overall system performance.

Purpose of the Study:

• To develop an advanced recognition approach for dynamic visual container goods in intelligent retail settings.
• To address the critical issues of hand occlusion and high product similarity that degrade recognition accuracy.

Main Methods:

• Utilized DarkNet53 as a backbone for feature extraction, incorporating semantic segmentation for occluded part localization and YOLOX for bounding box detection.
• Employed a generative adversarial network (GAN) with prior inference to restore and augment features of occluded regions.
• Integrated multi-scale spatial and effective channel attention mechanisms for fine-grained feature selection and a metric learning method based on von Mises-Fisher distribution for enhanced feature distinction.

Main Results:

• The proposed prior inference method showed improvements in peak signal-to-noise ratio (0.7743) and structural similarity (0.0183) compared to other models.
• Mean Average Precision (mAP) for recognition accuracy increased by 1.2%, with an overall recognition accuracy improvement of 2.82% over optimal models.
• The approach effectively handled goods occlusion caused by hands and distinguished between highly similar products.

Conclusions:

• The developed generative adversarial network combined with prior inference effectively resolves key challenges in intelligent retail product recognition.
• The method significantly enhances fine-grained feature distinction, leading to higher accuracy for occluded and similar goods.
• This research offers a robust solution with good application prospects for improving accuracy in intelligent retail commodity recognition.