Abstract
In the rapidly evolving landscape of smart retail, accurate commodity detection is critical for automated vending and inventory management. However, performance is often degraded by occlusion from shoppers' hands or adjacent products. This paper presents YOLO-R, an optimized object detection framework designed to mitigate these challenges. By integrating semantic segmentation for location priors, a lightweight Bi-directional Feature Pyramid Network (BiFPN) for multi-scale fusion, and a novel entropy-guided attention mechanism, YOLO-R effectively isolates unobstructed features. Experimental results on a custom retail dataset demonstrate a mean Average Precision (mAP) of 97.80% at 22.72 FPS, significantly outperforming standard YOLOv5 and YOLOv7 architectures in complex, occluded environments.
Introduction
The rise of "grab-and-go" retail containers necessitates robust computer vision systems capable of tracking items in real-time. A primary bottleneck in these systems is occlusion. When a customer reaches for a product, their hand frequently blocks a significant portion of the item, leading to missed detections or false positives by traditional convolutional neural networks (CNNs). Standard single-stage detectors often struggle to extract reliable features from these partially hidden objects. The goal of this research is to develop a high-speed detector that can dynamically prioritize visible features while ignoring occluding elements, ensuring high accuracy in dense retail displays.
Methodology: The YOLO-R Architecture
The authors propose YOLO-R, a specialized modification of the YOLOv5 architecture tailored for the unique constraints of retail environments. The pipeline consists of several key innovations:
1. Semantic Segmentation Pre-processing
To provide the detector with strong spatial priors, the system first employs an improved U-Net architecture enhanced with an Efficient Channel Attention (ECA) module. This stage performs semantic segmentation to distinguish the commodity from the background and the shopper's hand, effectively "cleaning" the input for the subsequent detection stages.
2. Entropy-Guided Attention (SENet)
The core innovation of YOLO-R is the injection of "feature entropy" into a Squeeze-and-Excitation (SENet) attention mechanism. By quantifying the effective characteristics of a commodity through entropy, the network can dynamically adjust weights. This allows the model to focus heavily on the discriminative, unobstructed pixels of a product while suppressing the influence of occluding objects like hands or shelf edges.
3. Lightweight BiFPN and Feature Fusion
To capture multi-scale details without excessive computational overhead, the standard feature pyramid is replaced with a lightweight Bi-directional Feature Pyramid Network (BiFPN). This component utilizes Content-Aware ReAssembly of FEatures (CARAFE) upsampling operators to better fuse high-resolution spatial information with low-resolution semantic data.
4. Dsoft-IOU Non-Maximum Suppression
In crowded retail shelves, items are often packed tightly, causing significant bounding box overlap. Standard Non-Maximum Suppression (NMS) often erroneously deletes valid boxes in these scenarios. The authors implement Dsoft-IOU, which incorporates the Euclidean distance between bounding box centers into the filtering logic, preserving accurate detections in dense clusters.
Experimental Results
Datasets
The model was validated using the standard VOC2007 dataset for benchmarking and a Self-Made Smart Retail Container Dataset. The latter contains 4,270 images of 21 commodity types (bagged, bottled, and canned) specifically curated to include varying degrees of occlusion, from slight to severe.
Performance Metrics
YOLO-R achieved a 97.80% mAP and a recall of 99.93% on the custom retail dataset. Crucially, it maintained a real-time processing speed of 22.72 FPS.
| Model | mAP (%) | Speed (FPS) |
|---|---|---|
| YOLOv5 (Baseline) | 97.40 | 18.89 |
| YOLOv7 | 96.50 | 20.24 |
| YOLO-R (Ours) | 97.80 | 22.72 |
Robustness to Occlusion
The model's primary strength lies in handling severe occlusion. In tests involving "Spring" water bottles under heavy blockage, YOLO-R maintained an accuracy of 68.88%, whereas the baseline YOLOv5 model dropped significantly to 52.75%. Ablation studies confirmed that the entropy-guided SENet alone contributed a ~0.4% boost in mAP while simultaneously increasing processing speed by over 2 FPS compared to standard SE implementations.
Conclusion and Future Work
YOLO-R demonstrates that a highly optimized single-stage detector can surpass more complex architectures in specialized tasks like retail commodity detection. By leveraging entropy to guide attention, the model effectively "sees through" occlusions. However, the authors note a susceptibility to high image noise; at very low Signal-to-Noise Ratios (SNR), detection performance degrades sharply. Future research will focus on improving the model's anti-interference capabilities and generalizability across diverse lighting and environmental conditions.
Summary updated