Unsupervised Feature Learning Via Spectral Clustering of Multidimensional Patches for Remotely Sensed Scene Classification

Abstract

Scene classification plays an important role in the interpretation of remotely sensed high-resolution imagery. However, the performance of scene classification strongly relies on the discriminative power of feature representation, which is generally hand-engineered and requires a huge amount of domain-expert knowledge as well as time-consuming hand tuning. Recently, unsupervised feature learning (UFL) provides an alternative way to automatically learn discriminative feature representation from images. However, the performances achieved by conventional UFL methods are not comparable to the state-ofthe-art, mainly due to the neglect of locally substantial image structures. This paper presents an improved UFL algorithm based on spectral clustering, named UFL-SC, which cannot only adaptively learn good local feature representations but also discover intrinsic structures of local image patches. In contrast to the standard UFL pipeline, UFL-SC first maps the original image patches into a low-dimensional and intrinsic feature space by linear manifold analysis techniques, and then learns a dictionary (e.g., using K-means clustering) on the patch manifold for feature encoding. To generate a feature representation for each local patch, an explicit parameterized feature encoding method, i.e., triangle encoding, is applied with the learned dictionary on the same patch manifold. The holistic feature representation of image scenes is finally obtained by building a bag-of-visual-words (BOW) model of the encoded local features. Experiments demonstrate that the proposed UFL-SC algorithm can extract efficient local features for image scenes and show comparable performance to the state-of-the-art approach on open scene classification benchmark.

Fast Binary Coding for the Scene Classification of High-Resolution Remote Sensing Imagery

“…Although COPD [12], which is an object-oriented scene classification framework, presents better performance than ours, it contains very complex pre-training to discover discriminative visual parts from images in the aspect of computational efficiency. The UFL-SC [40], which outperforms our method, also encompasses time-consuming manifold learning, nonlinear dictionary learning and a feature encoding stage. In other words, the proposed FBC is a compromise method that achieves a balance in accuracy and efficiency.…”

“…In [39], neural networks are used to train a set of feature extractors, with some techniques to reduce overfitting in the feature learning stage. Hu et al [40] propose an improved UFL pipeline where both learning model parameters and encoding features are performed on a low-dimensional manifold. It is worth mentioning that the latter two methods are free of any low-level hand-crafted features.…”

“…BOW [7] 71.86 SPM [30] 74 SCK [32] 72.52 SPCK++ [7] 77.38 SC + pooling [6] 81.67 ± 1.23 SG+ UFL [39] 82.72 ± 1.18 NFNC [71] 87.67 UFL-SC [40] 90.26 ± 1.51 COPD [12] 91.33 ± 1.11…”

The Fisher Kernel Coding Framework for High Spatial Resolution Scene Classification

“…The BOVW model, SPM [32], LDA [9], and LDA with a hybrid strategy (P-LDA) [30] were employed as the comparison methods, where the classifier of BOVW was SVM with a radial basis function (RBF) kernel. For the UCM dataset, the accuracies published in the previous works [15][16][17]19,22,[33][34][35] are also reported. …”

“…To bridge the semantic gap, scene classification methods based on the bag-of-visual-words (BOVW) model [13][14][15][16][17][18], part detectors [19,20], and neural networks [21][22][23] have been proposed, among which the BOVW model is one of the most popular approaches. In scene classification based on the BOVW model, the low-level features are extracted from the image by a local feature extraction method, e.g., mean/standard deviation statistics [9], the gray-level co-occurrence matrix [24], or scale invariant feature transform [25], and the low-level features are then assigned to their closest visual words in a "visual vocabulary", which is a codebook learned from a large set of local low-level features with k-means clustering.…”