add new cosine method

[MrTuanDao]

This update enhances the method for calculating cosine similarity to better handle low-quality images from miners. The previous approach relied on a CLIP model with a 384x384 resolution, which struggled to effectively evaluate low-quality issues in larger images.

Key Changes:

1. New Quality Score Method:

   • Images from miners and validators are cropped into several regions.
   • The top 10 regions with the highest pixel similarity are selected.
   • The mean of these top 10 regions' cosine similarity scores is used as the final quality score.

2. Old Method for Overall Similarity:

   • The original method is retained for calculating overall similarity between images.

3. Adjusted Weighting:

   • The weight of the original method has been adjusted to prioritize overall similarity, while incorporating the new quality score to refine evaluations.

This combination ensures a more robust assessment of image quality and similarity, particularly for identifying and mitigating low-quality miner images.