Abstract: The image target detection technology of the tank fire control system is an important basis for realizing the autonomous attack of the tank fire control system.In view of the image target detection task of tank fire control system,we establish a tank fire control system image dataset named TKHK which considers different target types,scales,deformation,light and shade,meteorological and other conditions,and can provide the basis for evaluation of different target detection methods.We propose a multi-scale deformation target detection method:Firstly,we use the ResNet-101-deformable convolution network and the deformable ROI pooling to improve the detection ability of the deformation target;then,we extract the proposed region on the convolutional feature maps with different resolutions,and we fuse the feature by adaptive feature fusion mechanism in the detection sub-network to improve the detection effect of multi-scale targets;finally,we combined with multiple design and training methods,such as online hard example mining,Soft-NMS and multi-scale training.Our method achieves better detection results on TKHK,and can better meet the actual application needs of the equipment.