Biological Vision-inspired Feature Extraction And Object Detection For High Resolution Remote Sensing Images

With the continuous development and progress of remote sensing technology, the spatial resolution of remote sensing images has been continuously improved, and the objects in images are clearly visible with rich details. The initially used feature extraction and recognition methods for mid- and low- resolution remote sensing images have been unable to meet the application requirements of high resolution remote sensing images. This brings unprecedented challenges to the automatic and semi-automatic classification and recognition of high resolution remote sensing images. Human visual system, with the outstanding ability of cognition and recognition, can easily identify various objects in the complex high-resolution remote sensing images. Inspired by this, feature extraction and object detection methods that simulate the information processing mechanisms in human visual system are studied in this paper. This study will give great technical support to the automatic or semi-automatic classification and recognition of high resolution remote sensing images.Based on the in-depth understanding of information processing mechanism of human visual system, especially the visual information processing mechanism of the ventral pathway and visual attention, the hierarchies of perception and cognition of human visual system are corresponded to the low-level visual features, mid-level features and high-level abstract semantic features. Our research concerns sparse texture model, hierarchical visual cortex model and visual attention model, and this paper focuses on the texture feature extraction and object hierarchical visual feature extraction for the land use classification of high resolution remote sensing images. For the recognition of given objects in high resolution remote sensing images, the object detection method based on visual attention is also studied.The innovations of this paper are as follows:(1) The Orientation Pooling based on Sparse Representation method(OPSR) is proposed to extract sparse and invariant texture feature. In OPSR, the over-complete dictionary for sparse representation is taken as a population of neurons in the cerebral cortex, and each atom in the dictionary is like a simple cell in the visual cortex that can respond to a stimulus with a specific orientation. Then, each atom is rotated at several different angles, which generates the neurons that can respond to stimuli with different orientations. All atoms and their rotated ones together constitute an extended dictionary. The sparse decomposition of an image based on the extended dictionary can be taken as the sparse responses of neurons to given stimuli with different orientations. Simulating the pooling of complex cells in visual cortex of human visual system, all responses from an atom(including its rotated ones) are pooled to extract the rotation invariant features. The experiments on texture images show that the proposed method can extract rotation invariant features, which conduces to object classification and recognition of high resolution remote sensing images.(2) The enhanced standard model of visual cortex(enhanced ST model) is proposed to extract features of remote sensing images. This model not only simulates the sparse coding strategy in human visual system(HVS), but also uses the trained over-complete dictionary as prototype dictionary to substitute for the original random constructed prototype dictionary. Then, the C2 features are obtained further based on the new prototype dictionary. The way to obtain the prototype dictionary by learning is more consistent with human cognitive processes. The results of two-class and multi-class classification experiments show that the proposed model is more stable, effective and efficient comparing with the standard model of visual cortex, and it also has a high distinguishing ability to natural objects with regular textures.(3) Simulating the mechanism of visual selective attention in human visual system, the multi-cues integrated visual attention model based on Bayesian inference is proposed. In this model, the visual attention is taken as Bayesian inference process which can deal with the uncertainty of visual cues. It uses the bottom up visual cues and top-down priori information, and all cues and priori information are associated with location information and inputted into the Bayesian network. Then, the Bayesian inference is implemented to simulate human visual selective attention. The experiments show that the proposed model achieves good performance in object detection for high resolution remote sensing images with complex backgrounds.