Grid Anchor Based Image Cropping: A New Benchmark and An Efficient Model.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Existing image cropping methods lack flexibility and robust evaluation metrics.
• Current databases offer limited ground truth, failing to capture cropping's non-unique nature.

Purpose of the Study:

• To develop a more effective and efficient image cropping formulation and benchmark.
• To address limitations in existing cropping databases and evaluation metrics.

Main Methods:

• A grid anchor-based formulation considering local redundancy, content preservation, and aspect ratio.
• Construction of a grid anchor-based cropping benchmark with comprehensive annotations.
• Design of a lightweight, multi-scale cropping model focusing on regions of interest and discard.

Main Results:

• Reduced search space for candidate crops from millions to under ninety.
• Developed a benchmark with reliable evaluation metrics for image cropping.
• Achieved high-speed performance (200 FPS on GPU, 12 FPS on CPU) with a lightweight model (<2.5M parameters).

Conclusions:

• The grid anchor formulation offers a more practical and efficient approach to image cropping.
• The new benchmark and metrics provide a more reliable evaluation of cropping models.
• The proposed lightweight model delivers robust and visually pleasing results efficiently.