Abstract: The occlusion interference of smoke screen and cloud can change the target characteristics and cause the target identification errors when an infrared imaging system detects and identifies a target. The targeted processing during the identification process can greatly reduce the identification false alarm rate and improve the anti-interference ability of identification by performing the positioning and type judgment of occlusion interference area. To this end, an infrared image thick cloud and smoke screen occlusion interference detection method based on improved deep learning single channel detector is proposed. In the proposed method, the multi-scale prediction is realized by multiplexing and merging the multi-layer features of network, and the detection precision is inceased by using the dynamic anchor frame module to improve the anchor frame mechanism. The detection speed is inceased by merging the convolutional layer and the batch normalization layer in the network, and the classification ability of network for the obstruction is improved by introducing the central loss function to optimize the classification function. An infrared sample augmentation method is proposed for network training, which effectively expands the data volume and solves the problem of difficult acquisition of infrared image training samples. The experimental results show that the proposed method is used to improve the detection accuracy by 3.7% at the same speed, which effectively solves the problem of weak anti-interference ability of infrared automatic target recognition system in complex environment. Key

Key words: infraredimage, occlusioninterferenceidentification, convolutionalneuralnetwork, sampleaugmentation