Study On Underwater Object Automatic Recognition Based On Aerial Visible Remote Sensing

To satisfy the need of detecting ocean underwater target by aerial visible light remote sensing, this paper develops an underwater target automatic recognition system prototype. The system prototype collects visible light information by digital camera, processes images by computer real-timely, obtains underwater target information quickly, and reduces error of traditional method. The traditional underwater target monitoring is mainly accomplished by observing ocean color abnormity from telescope. Sometimes, it is very difficult to detect underwater target and computer their parameter such as type, shape and so on for ocean phenomena disturbance. Because of absorbing and scattering of the water, a few fraction of underwater target information is received by the remoter sensor. Otherwise, ocean random disturbance,sun flares and cloud shadows induce target information weak while induce noisy information robust. The difficult problem of this paper is to recognize weak information underwater target on the condition of robust information noisy. The primary achievement and innovation are as follow.1) This paper proposes the optimal color space selection for underwater object recognition based on water spectrum feature and object spectrum feature. It is different between spectrum reflectivity of underwater object and water, while it is nearly same between spectrum reflectivity of underwater object and object model. Thus it is reasonable to study object model spectrum feature to simulate underwater object spectrum feature. A lot of experiments are done to measure spectrum reflectivity of object model and water. The spectrum reflectivity curves are analyzed to find wave band with the most reflectivity difference between underwater object and water and achieve the optimal color space. The validity of selected color space is proved from physics mechanism and experiment images.2) This paper proposes a fast algorithm for two-dimensional otsu adaptive threshold algorithm. As a classical image segmentation method, otsu adaptive threshold algorithm has applied widely in image processing. The application of the two-dimensional otsu threshold algorithm based on the otsu threshold algorithm has been restricted for the long-paying computation. This paper gives a fast algorithm for two-dimensional otsu adaptive threshold algorithm that overcomes the disadvantage of high computational complexity. This fast algorithm gets rid of redundance computation and yields a look-up table by iteration. The computational time of the fast method is not only far less than that of the source two-dimensional one, but also yields the same threshold and the original method.3) This paper proposes the high illumination area threshold algorithm based on fisher separation criterion and two dimensional otsu threshold. First, the whole image is divided into sub-images with the same size. Second, the sub-image is segmented into two classifications by two dimensional otsu threshold. The segmentation is validated by the fisher criterion to determine whether the sub-image threshold is empty. Finally, the sub-image with empty threshold is set to average threshold of its neighbor sub-images. From experiments, the method can availably segment high illumination area of sprays and sun flares to eliminate cloud shadows and illumination variety.4) This paper proposes an edge restoration algorithm of underwater target based on deformable templates. Because of sun flares, noisy and so on, it is mostly impossible to directly extract the edge contour of underwater target. However, it is necessary to confirm position and size of the underwater target in practical application. Thus we define the energy function of deformable templates anew according to underwater target shape prior information. Gradient descending, simulate annealing, genetic algorithm and immunity algorithm are used to solve the minimum value of energy function. From a lot of experiments, the immunity algorithm precedes other three algorithms.5) This paper proposes a wake detection method for the water wake of airphotoes using independent component analysis of Gabor features and based on two-dimensional principal component analysis (2DPCA) of directional polar Fourier spectrum. Based on a lot of texture analysis algorithm, according to the strong direction performance of strip texture, this paper proposes the weighted covariance matrix to obtain the strip texture orientation more precisely. Also texture analysis algorithms based on principal component analysis, independent component analysis of Gabor features and based on two-dimensional principal component analysis of Directional Polar Fourier Spectrum. Statistical learning theory and support vector machine are studied thoroughly to be the classification of texture features. From lots of experiment results, it is proved that the proposed algorithm can extract wake texture of underwater target precisely.To validate our algorithm of detecting underwater targets in aerial visible light remote sensing images, many experiment images are done to achieve very high correct percent. However, there are many noisy in the ocean such as random disturbance, sun flares, cloud shadows, and weak underwater target information. So more experiments are needed to perfect an underwater target automatic recognition system prototype and practice the prototype.