A robust framework combined saliency detection and image recognition for garbage classification.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Current garbage classification models often lack robustness due to simple backgrounds in training datasets like Trashnet.
• Real-world scenarios present complex backgrounds that degrade the performance of existing garbage classification systems.

Purpose of the Study:

• To develop a more robust garbage classification framework capable of handling complex backgrounds.
• To improve the generalization performance and accuracy of automated garbage identification systems.

Main Methods:

• A framework integrating saliency detection (Salinet) and image classification (Inception V3) was proposed.
• Garbage segmentation was achieved by identifying the target area using saliency detection and creating a bounding rectangle.
• Training data was augmented by fusing original Trashnet images with complex backgrounds from other datasets.

Main Results:

• The proposed framework demonstrated significant accuracy improvements (0.50%–15.79%) on datasets with complex backgrounds.
• Achieved a 4.80% accuracy gain on a collected real-world dataset, outperforming existing methods.
• The method proved more robust to noise and complex environmental conditions.

Conclusions:

• The combined saliency detection and classification approach enhances garbage classification robustness in complex environments.
• This framework offers a practical solution for automatic garbage classification in smart trash cans.
• The study highlights the importance of diverse and complex background data for training effective computer vision models.