Detection of small objects in image data is fundamental to many image processing applications. Spatial domain detection of small objects is the key process of the detect before track (DBT) method for moving object detection. Most of the existing spatial detection methods are filter-based ones. We present a novel and efficient approach to spatial detection of small objects in image data, which combines the local signal-to-noise ratio (SNR) characteristic and appearance characteristic of small objects. In such a detection scheme, the nonlinear principal component analysis (NLPCA) neural network (NN) is used for modeling the appearance of small objects and constructing a saliency measure function. Based on this function and the feature vector extracted at each pixel position using the principal component analysis (PCA) technique, a small object saliency map is formed by lexicographically scanning the input image, then the saliency map is thresholded to obtain the intermediate object location map. We also treat such a saliency map as a spatially filtered result of the input image. Compared to several filter-based detection methods, experiments show that the proposed algorithm outperforms these methods.