Infrared target recognition has two problems:the first one is the difficulty to extract invariant feature owing to the blur infrared images;The second one is that the infrared image has little information and it is hard to be represented in an effective descriptor.To solve the problems,a detection method of affine invariant triangles based on corners in curvature scale space (CSS) is presented,and the triangular region is represented by using multi-scale auto-convolution (MSA) whose parameters are optimized.At last,infrared target recognition can be realized by matching features.Experimental results show that recognition effect of this method for infrared targets is better than other methods in many kinds of image transformations.