AGMG-Net: Leveraging multiscale and fine-grained features for improved cargo recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Security systems require high accuracy in cargo identification to prevent economic losses.
• Current cargo identification methods struggle with limited data, feature utilization, and subtle class differences, hindering accuracy.
• Achieving 99.99% recognition accuracy remains a significant challenge in cargo identification modules.

Purpose of the Study:

• To develop an advanced model for accurate cargo identification.
• To address the limitations of existing methods by improving feature extraction and fusion.
• To enhance the reliability and effectiveness of security systems through precise cargo recognition.

Main Methods:

• A novel cargo identification dataset, "Cargo", was created using industrial cameras.
• An Attention-guided Multi-granularity feature fusion model (AGMG-Net) was proposed.
• AGMG-Net employs two branch networks for coarse- and fine-grained feature extraction, fused via a Concat manner.
• Attention-guided Multi-stage Attention Accumulation (AMAA) and Multi-region Optimal Selection Based on Confidence (MOSBC) modules were introduced for localization and cropping.

Main Results:

• The AGMG-Net model achieved high recognition rates: 99.58% on the "Cargo" dataset, 92.73% on "Flower", and 88.57% on "Butterfly20".
• The proposed model outperformed state-of-the-art methods in cargo identification accuracy.
• Experimental results demonstrate the model's effectiveness in accurately identifying diverse cargo categories.

Conclusions:

• The developed AGMG-Net model significantly improves cargo identification accuracy.
• This research offers a valuable solution for enhancing the capabilities of security systems.
• The proposed method provides accurate cargo categorization, contributing to improved security and reduced economic damage.