JigsawNet: Shredded Image Reassembly Using Convolutional Neural Network and Loop-Based Composition

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Image reassembly from shredded fragments is a challenging problem.
• Existing methods often struggle with complex puzzles due to reliance on handcrafted features and greedy algorithms.
• Reliable pairwise matching and global composition are critical for accurate image reconstruction.

Purpose of the Study:

• To develop a novel algorithm for reassembling arbitrarily shredded images.
• To improve the accuracy and efficiency of image reassembly pipelines.
• To address limitations in existing local matching and global composition stages.

Main Methods:

• A deep convolutional neural network (CNN) was developed to detect pairwise stitching compatibility, pruning incorrect matches.
• CNN calculations were optimized by transferring them to the stitching region with a boost training strategy.
• Two new loop closure-based searching algorithms were proposed for the global composition stage, replacing greedy strategies.

Main Results:

• The proposed algorithm significantly outperforms existing methods in solving various image reassembly puzzles.
• The method demonstrates superior performance on challenging puzzles with a large number of fragment pieces.
• Experimental results validate the effectiveness of the CNN-based matching and loop closure-based composition.

Conclusions:

• The novel algorithm offers a significant advancement in image reassembly technology.
• The integration of deep learning for local matching and advanced search for global composition provides robust solutions.
• This approach is particularly effective for complex and large-scale image reconstruction tasks.