Detection of Infrared Small Targets Using Feature Fusion Convolutional Network

This paper proposes the feature extraction backbone network MNET that is specifically designed for the detection of infrared small targets. The overall network uses three down-sampling operations to adjust the size of the feature map, while preserving sufficient physical characteristics of the small infrared target to be used in the detection. In a next step, the dense connection is used to save the output of each layer of the network in the front channel of the feature map, to better integrate the location information of the shallow network and the semantic information of the deep network. In this way accurate network positioning and classification effects are achieved. As a last step, we introduce a feature attention mechanism to obtain the importance of each feature channel, and to enhance useful features according to their degree of importance. In this way we achieve an adaptive calibration of the feature channels. In order to train the proposed detection network MNET from scratch, the single-phase detection algorithm YOLO is adopted for the detection part. To verify the effectiveness of the proposed method, we captured images and created an infrared small target dataset. The experimental results show that MNET can accurately detect targets of
$2\times2$
pixels size in infrared images of
$640\times512$
pixels at a processing speed of up to 105 frames per second. MNET meets real-time requirements while providing high quality detection accuracy.