To solve the problems of the existing CNN-based methods,such as feature information loss,serious clutter information interference,ignoring the correlation between different scale features,and requiring a large number of training samples,a tank detection method based on Siamese Feature-Guided Multi-Scale Network (SFGMSN) is proposed.In SFGMSN,an improved Inception module is designed to extract the multi-scale features of the tank target image and carry out feature fusion to better recover the fine segmentation information of the tank target.To improve the feature perception ability of the target region and suppress background interference,a Local Channel Attention Mechanism (LCA-M) is designed to obtain more accurate detection results.Finally,ameta-learner is used to detect tank targets.SFGMSN makes full use of the advantages of multi-scale convolution,dilated convolution,Siamese network,local channel attention mechanism and the meta-learner.It can solve the problems that the traditional CNN model over-relies on a large number of training samples and may have low accuracy and poor generalization under the condition of small sample size.The experimental results on the tank image dataset show that the proposed method is effective in tank detection,with an average detection accuracy of 90.12%.It can realize tank detection in complex scenes and has good robustness to low-resolution tank images.