Detection of seals in remote sensing images using features extracted from deep convolutional neural networks

In this paper, we propose an algorithm for automatic detection of seals in aerial remote sensing images using features extracted from a pre-trained deep convolutional neural network (CNN). The method consists of three stages: (i) Detection of potential objects, (ii) feature extraction and (iii) classification of potential objects. The first stage is application dependent, with the aim of detecting all seal pups in the image, with the expense of detecting a large amount of false objects. The second stage extracts generic image features from a local image corresponding to each potential seal detected in the first stage using a CNN trained on the ImageNet database. In the third stage we apply a linear support vector machine to classify the feature vectors extracted in the second stage. The proposed method was demonstrated to an aerial image that contains 84 pups and 128 adult harp seals, and the results show that we are able to detect the seals with high accuracy (2.7% for the adults and 7.3% for the pups). We conclude that deep CNNs trained on the ImageNet database are well suited as a feature extraction module, and using a simple linear SVM, we were able to separate seals from other objects with very high accuracy. We believe that this methodology may be applied to other remote sensing object recognition tasks.