Referring to the current use of sparse representation algorithms to extract camouflaged targets from hyperspectral images, the selection of the background dictionary is affected by the “same spectrum of different objects” of the hidden targets, resulting in the inability to detect camouflaged targets accurately. In this paper, we take the grassland camouflage net and desert camouflage net as the research objects, collect the visible-near-infrared reflectance spectra of the camouflage net and airborne hyperspectral images respectively, and analyze the spectral characteristics of the background pixels and camouflage target pixels in the camouflage net and airborne images measured outdoors. Taking advantage of the fact that the spectra of the camouflage nets measured outdoors and the background in the airborne images are different and the possibility of neighboring image elements belonging to the same feature is high, the background dictionary selection method based on the constraints of Euclidean distances and image homogeneity features is proposed, and the sparse representation of the background dictionary is further utilized to identify the target. The results show that (1) in the wavelength range of 750~1 000 nm for grass camouflage, the reflectance of the background pixel spectrum in the image is higher than that of the camouflage net spectrum. For desert camouflage: in the range of 550～700 nm, the reflectance of the background pixel spectrum in the image is higher than that of the camouflage net spectrum. (2) By establishing spatial and spectral feature constraints with the maximum spectral Euclidean distance to the target image element and the highest homogeneity with neighboring image elements, 413 background image elements in the grass camouflage image and 507 background image elements in the desert camouflage image were selected as the background dictionary. (3) Based on the improved background dictionary selection method, the sparse representation algorithm is utilized to identify the camouflage targets, and the results can accurately discriminate the location and number of camouflage targets. The area under the curve (AUC) of the receiver operating characteristics for the detection of grass camouflage targets and desert camouflage targets reaches 0.96 and 0.98, respectively, indicating that the algorithm has good detection performance for both grass camouflage targets and desert camouflage targets.