A Fine-Grained Image Classification Approach for Dog Feces Using MC-SCMNet under Complex Backgrounds.

Area of Science:

• Computer Vision
• Machine Learning
• Veterinary Science

Background:

• Dog feces identification is challenging due to environmental degradation (weathering, sun exposure) and background disturbances (decaying wood, dirt).
• Subtle visual distinctions between different types of feces complicate accurate classification.
• Existing methods struggle with the fine-grained classification of dog feces in natural, complex environments.

Purpose of the Study:

• To propose a fine-grained image classification approach, MC-SCMNet, for robust dog feces identification in complex natural backgrounds.
• To develop novel modules (MADM and CLAM) to enhance feature extraction and suppress noise.
• To improve the accuracy and stability of dog feces recognition for potential application in canine health diagnostics.

Main Methods:

• Developed a Multi-scale Attention Down-sampling Module (MADM) to capture subtle fecal features.
• Introduced a Coordinate Location Attention Mechanism (CLAM) to mitigate interference from background disturbances.
• Constructed a new backbone network using SCM-Blocks (integrating MADM and CLAM) and Depthwise Separable Convolution (DSC) for parameter efficiency.

Main Results:

• MC-SCMNet achieved superior accuracy compared to other models on the self-built DFML dataset.
• The model attained an average identification accuracy of 88.27% and an F1 score of 88.91%.
• Demonstrated stable performance even in complex background conditions.

Conclusions:

• MC-SCMNet is highly appropriate for dog fecal identification tasks.
• The model's robustness in complex environments suggests potential applications in monitoring dog gastrointestinal health.
• Fine-grained image classification with specialized attention mechanisms offers a promising approach for environmental sample analysis.