Real-time Object Detection Model for Cigarette Brand Identification Based on an Improved Single-stage Regression...

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Automated cigarette inventory systems face challenges with visual recognition due to illumination variations, diverse box sizes, and occlusion.
• Accurate brand verification and detection of misplaced items are critical for inventory management.

Purpose of the Study:

• To develop an improved real-time detection model for enhanced accuracy in visual recognition tasks for cigarette inventory.
• To address limitations of existing models in handling complex visual conditions such as occlusion and varying lighting.

Main Methods:

• An adaptive downsampling module was integrated into the YOLO series backbone and neck networks to preserve feature details and reduce model size.
• An inverted efficient multi-scale attention module was employed to capture spatial context and improve accuracy for occluded or complex features.
• A dynamic detection head module was implemented for multi-scale object detection, enhancing positioning and classification.

Main Results:

• The proposed model achieved a mean Average Precision (mAP) of 97.9% on a custom cigarette box dataset.
• The model demonstrated a 0.9% improvement in mAP over the baseline, with a 28.78% reduction in parameters and 1.4 G reduction in floating-point operations.
• An inference speed of 38.5 frames per second (FPS) was achieved, meeting real-time industrial detection requirements.

Conclusions:

• The improved visual recognition model effectively meets the demands of real-time cigarette inventory detection.
• The model's enhancements in feature detail retention, spatial context capture, and multi-scale detection contribute to its superior performance.
• The developed model offers a lightweight, accurate, and efficient solution for automated visual inventory systems.