Target detection of hyperspectral image has always been a hot research topic, especially due to its important applications in military and civilian remote sensing. This paper employs the idea of classification and proposes a novel detection framework which incorporates dictionary learning and discriminative information. Due to the fact that target pixels lie in different subspace with background pixels, a novel detection model is proposed. In addition, a linear kernel is applied to project the image data into high-dimensional space, separating the target pixels and background pixels. Synthetic image and popular real hyperspectral image are used to evaluate our algorithm. Experimental results indicate that our proposed detector outperforms the traditional detection methods. © 2019 SPIE.