Research Of Ship Detection In Foggy Optical Satellite Image Based On Improved RPN+PCANet

Ship is a common transportation at sea,and the realization of automatic ship detection at sea has great significance to both civil and military applications.In recent years,with the gradual improvement of optical satellite image quality and the successful application of convolution neural network technology in the field of image classification and object recognition,ship recognition based on optical satellite image has gradually attracted the attention of many researchers.This paper mainly focuses on ship recognition in optical satellite images with foggy weather,and carry out a series of related research including foggy removal pre-processing for experimental image,ship image classification and ship detection.The main research contents are summarized as follows.(1)Defogging preprocessing of optical satellite images based on dark channel haze removal algorithm.Because optical satellite images are susceptible to weather,especially in foggy weather,the recognizable objects in the images will be blurred,thus reducing the accuracy of recognition.In this paper,after verifying that optical satellite images satisfy the dark channel priori rules,the fogging images are pre-processed to eliminate the influence of foggy environment based on the dark channel haze removal algorithm.(2)In ship image classification phase,because of the disadvantage of discontinuous recognition process in PCANet(Principal Component Analysis Network),end-to-end PCANet is proposed to improve the integrity of the network by adding spatial pyramid pooling and fully connected network.This paper compares the proposed end-to-end PCANet with other classification algorithms by different test datasets,and the experimental results show that the end-to-end PCANet not only improves classification accuracy,but also has good anti-noise ability.(3)In ship detection phase,this paper improves RPN(Region Proposal Network)by connecting different convolution feature maps.The final model,named improved RPN+PCANet,consists of two parts.The first part is RPN that is used for recommending all possible region,and the second part uses end-to-end PCANet as classification network to classify all possible region.The model is compared with common object recognition algorithms on different test datasets,and the experimental results show that the ship recognition algorithm proposed in this paper can effectively recognize ship targets in optical satellite images in foggy environment.