Dim and small target detection based on feature mapping neural networks

摘要

Dim and small target detection based on passive millimeter wave or infrared imaging is of great value in both security and military fields and has been studied extensively. The problems of weak distinction between small targets and backgrounds and of less extractable features of targets have always been a technical bottleneck for accurate detection of dim and small targets. For dim and small targets with few pixel-based features on complex and diverse backgrounds, we propose a high-precision detection algorithm based on feature mapping deep neural networks with a spindle network structure. Firstly, the features of low-dimension dim and small target blocks are mapped to a higher-dimensional space. An encoded neural network is then used to extract high-discriminant features to complete the background and target recognition. Background suppression and target enhancement is realized according to the intensity (the distinguished output of the network). Finally, a detection method based on the constant false alarm rate is used to detect dim and small targets. The experimental results show that, compared with several popular algorithms for millimeter-wave and infrared image detection in different scenarios, the proposed algorithm has a lower false alarm rate, higher detection accuracy and stronger robustness. Statistics for experiments on under various false alarm rates and signal-to-noise ratios show that the detection rate of the proposed method is about 15% higher than that of the compared algorithms. In experiments on real data, the detection rate of our algorithm is more than 25% higher than that of the suboptimal algorithm. (C) 2019 Elsevier Inc. All rights reserved.