研究目的
Investigating the effectiveness of deep learning in small target detection with controlled signal to background noise ratio (SNR).
研究成果
The proposed deep learning method for small target detection significantly outperforms traditional filtering methods. Training with samples of specific constant SNR (SNR≈1) yields the best performance. This approach not only benefits small target detection but could also be applicable to other areas requiring signal detection from noisy backgrounds.
研究不足
The study is limited by the dependency on the specific SNR for training samples and the computational resources required for deep learning models. The performance of deeper nets did not improve beyond 5 full connection layers.
1：Experimental Design and Method Selection:
The study proposes an end-to-end deep learning solution for small target detection, utilizing deep neural networks with convolution and pooling layers.
2：Sample Selection and Data Sources:
Random background parts are sampled from cloud-sky images, and random generated target spots are added to these backgrounds with controlled SNR to generate target samples.
3：List of Experimental Equipment and Materials:
A laptop computer with Intel Core i7-2630QM CPU and GTX650m Nvidia Graphic Card was used for training and tests conducted on MATLAB platform with matconvnet toolbox.
4：Experimental Procedures and Operational Workflow:
The process involves generating target samples with controlled SNR, training deep nets of different architectures, and evaluating their performance on a test image with randomly added small spot targets.
5：Data Analysis Methods:
Performance is compared using Small Area Signal-to-Noise ratio gain (SSNR Gain) and Large Area Signal-to-Noise ratio gain (LSNR gain).
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
