We propose a multilayer system to perform ice image retrieval. Ice images are typically texture-less, which adds difficulty in retrieving the images. To achieve high accuracy, high level local features are usually used in retrieving the images. However, most high level features contain high dimensionality that slows down the retrieval process. To overcome this problem, we divide the retrieval process into 3 steps. Each step filters out a large portion of images. As the features are constructed according to the ice image properties, one image can be quickly localized compared with the use of high-level features. The ice images are captured in Arctic, where the ice state changes dramatically due to the environmental and other influences. We build the first layer of the system on the utilization of color information and edges, as the color and the edges are the most critical characteristics of ice images. We divide the second layer into two sub-layers. The first sublayer is on the use of edge histogram. For the second sublayer, we detect salient points based on pixel values on the edge position and connect every adjacent points with straight lines. A new feature is built on the basis of distance scale of every adjacent salient points and the angles between connected lines. Our new feature is invariant to transformation, rotation and scaling. As the features in the first two layers are holistic features, the time performance is much better than high-level local features. The third layer is to apply Harris detector to find the correspondences between two features on a small set of filtered images. The experiments show that our system achieves good accuracy while maintaining much better time performance.