Ship recognition based on Hu invariant moments and convolutional neural network for video surveillance

To solve the issue with automatic recognition of ship images in video surveillance system, this study proposes a ship recognition approach based on Hu invariant moments and Convolutional Neural Network (CNN). Ship image is firstly denoised using improved threshold function of Wavelet Transform (WT) and then segmented via iterative auto-threshold segmentation algorithm to extract the ship area in the image. After that improved CNN is applied to further extract features of ship images. At the same time the image is divided into sub-images horizontally where Hu invariant moments is extracted. The Hu invariant moments of each sub-image are concatenated into a composite vector as the Hu invariant moments of the whole image, which are integrated with spatial location information. The last step is to fuse the CNN features and Hu invariant moments to obtain discriminative feature representation. Hu invariant moments which are invariant to translation, rotation and scaling are used to supplement the shape and contour information of the ship images. And Softmax function is applied to automatically recognize ship images in the output layer. Two fusion methods have been applied to verify the effectiveness of the proposed approach based on feature extraction at different levels. Experimental results show that the first fusion method achieves highest recognition accuracy in self-built dataset and visible and infrared spectrums (VAIS) dataset, up to 98.28% and 92.80% respectively. Recognition accuracy of the second fusion method is also higher than existing methods. Moreover, results obtained from F1-score and confusion matrix further validate the effectiveness of the proposed approach.