研究目的
To propose a novel multi-level features convolutional neural network (MLFCNN) architecture for generating a single clean image by fusing multiple input images, aiming to improve fusion quality over existing methods.
研究成果
The proposed MLFCNN method effectively preserves original image information, handles slight object motion, and outperforms state-of-the-art methods in both subjective and objective evaluations. It demonstrates robustness and high fusion quality, with potential for further improvement using full convolution networks.
研究不足
The method relies on training with image patches, which may lead to inaccuracies at boundaries between focused and defocused regions. Future work could explore full convolution networks for pixel-level prediction to address this.
1：Experimental Design and Method Selection:
The study designs a MLFCNN with multi-level features and 1x1 convolutions to reduce redundancy. It involves pre-training the network on a dataset, using it to generate focus maps from source images, applying post-processing (morphological operations and Gaussian filtering) to obtain decision maps, and fusing images via weighted-sum.
2：Sample Selection and Data Sources:
Training uses the CIFAR-10 dataset to generate 60 million image patch pairs (16x16 size) with data augmentation (horizontal flip, vertical flip, random rotation). Testing uses 10 sets of multi-focus images, including 5 from the 'Lytro' dataset and others collected online.
3：List of Experimental Equipment and Materials:
No specific physical equipment mentioned; computational resources and software (e.g., TensorFlow) are implied for training and testing the neural network.
4：Experimental Procedures and Operational Workflow:
Steps include converting images to grayscale, feeding them into the pre-trained MLFCNN to get focus maps, thresholding to binary maps, applying morphological operations and Gaussian filtering for refinement, and fusing using weighted-sum.
5：Data Analysis Methods:
Evaluation uses mutual information (MI), gradient-based fusion metrics (Q_AB/F), and structure similarity-based metrics (Q_S) for objective assessment, with comparisons to seven other fusion methods.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
