Hyperspectral images provide significant information about spectral characteristics of the objects in the scene. Anomaly detection is to detect targets whose signatures are spectrally distinct from their surroundings without a priori knowledge.Generally

the prior knowledge of the targets is unknown

and therefore the anomaly detection is very impor- tant and challenging. Constrained energy minimization（CEM）proposed by Chang is a target detection algorithm.It constrains the desired target signature with a specific gain while minimizing the output energy of the whole signatures in the images.The problem is that it needs the signature of desired target as a priori knowledge.Therefore it is not appropriate for anomaly detection. Based on the CEM algorithm

we presents a local energy maximal division（LEMD）algorithm to achieve anomaly detec- tion in this paper. There are four innovations in the algorithm.Firstly

the paper proposes local energy maximal division（LEMD）algo- rithm.It localizes the detection area with a sliding window.The central pixel of the sliding window is regarded as the test- ed pixel

and the other pixels in the window are regarded as background.The idea of the algorithm is to use the maximal projection energy ratio between the tested pixel and the background to measure anomalous degree of the tested pixel. Secondly

the optimal projection vector is composed of the signature of the tested pixel and the inverse of correlation matrix estimated by background pixels.If there are anomalies in the local background

the correlation matrix will be contaminated and the tested anomaly may be missed.To solve the problem

the anomalies in the local background should be found and removed

and the employed method is based on the projection distribution of the local window.Pixel in the background is defined as anomaly if the following condition is satisfied:The projection output of the pixel is similar to that of the tested pixel and much different from that of the most background pixels.Thirdly

false alarms are generally existed in anomaly detection

and the multi-random-window voting method is proposed to control false alarm rate.The method acts on the potential anomalies produced by threshold segmentation.It chooses some local windows selected randomly in the images

and determines the potential anomalies whose anomalous degree is larger than certain threshold in most of the random win- dows as real anomalies.Experimental data are hyperspectral images with 210 bands produced by hyperspectral digital image collection experiment（HYDICE）.The comparative experiments between the LEMD algorithm and the well-known RX algorithm indicate that LEMD algorithm is robust and performs better under complex unknown background.