A Multi-Domain Collaborative Transfer Learning Method with Multi-Scale Repeated Attention Mechanism for Underwater Side-Scan Sonar Image Classification

Due to the strong speckle noise caused by the seabed reverberation which makes it difficult to extract discriminating and noiseless features of a target, recognition and classification of underwater targets using side-scan sonar (SSS) images is a big challenge. Moreover, unlike classification of optical images which can use a large dataset to train the classifier, classification of SSS images usually has to exploit a very small dataset for training, which may cause classifier overfitting. Compared with traditional feature extraction methods using descriptors—such as Haar, SIFT, and LBP—deep learning-based methods are more powerful in capturing discriminating features. After training on a large optical dataset, e.g., ImageNet, direct fine-tuning method brings improvement to the sonar image classification using a small-size SSS image dataset. However, due to the different statistical characteristics between optical images and sonar images, transfer learning methods—e.g., fine-tuning—lack cross-domain adaptability, and therefore cannot achieve very satisfactory results. In this paper, a multi-domain collaborative transfer learning (MDCTL) method with multi-scale repeated attention mechanism (MSRAM) is proposed for improving the accuracy of underwater sonar image classification. In the MDCTL method, low-level characteristic similarity between SSS images and synthetic aperture radar (SAR) images, and high-level representation similarity between SSS images and optical images are used together to enhance the feature extraction ability of the deep learning model. Using different characteristics of multi-domain data to efficiently capture useful features for the sonar image classification, MDCTL offers a new way for transfer learning. MSRAM is used to effectively combine multi-scale features to make the proposed model pay more attention to the shape details of the target excluding the noise. Experimental results of classification show that, in using multi-domain data sets, the proposed method is more stable with an overall accuracy of 99.21%, bringing an improvement of 4.54% compared with the fine-tuned VGG19. Results given by diverse visualization methods also demonstrate that the method is more powerful in feature representation by using the MDCTL and MSRAM.

Physics-Based Modelling and Simulation of Multibeam Echosounder Perception for Autonomous Underwater Manipulation

One of the key distinguishing aspects of underwater manipulation tasks is the perception challenges of the ocean environment, including turbidity, backscatter, and lighting effects. Consequently, underwater perception often relies on sonar-based measurements to estimate the vehicle’s state and surroundings, either standalone or in concert with other sensing modalities, to support the perception necessary to plan and control manipulation tasks. Simulation of the multibeam echosounder, while not a substitute for in-water testing, is a critical capability for developing manipulation strategies in the complex and variable ocean environment. Although several approaches exist in the literature to simulate synthetic sonar images, the methods in the robotics community typically use image processing and video rendering software to comply with real-time execution requirements. In addition to a lack of physics-based interaction model between sound and the scene of interest, several basic properties are absent in these rendered sonar images–notably the coherent imaging system and coherent speckle that cause distortion of the object geometry in the sonar image. To address this deficiency, we present a physics-based multibeam echosounder simulation method to capture these fundamental aspects of sonar perception. A point-based scattering model is implemented to calculate the acoustic interaction between the target and the environment. This is a simplified representation of target scattering but can produce realistic coherent image speckle and the correct point spread function. The results demonstrate that this multibeam echosounder simulator generates qualitatively realistic images with high efficiency to provide the sonar image and the physical time series signal data. This synthetic sonar data is a key enabler for developing, testing, and evaluating autonomous underwater manipulation strategies that use sonar as a component of perception.

Side-Scan Sonar Image Classification Based on Style Transfer and Pre-Trained Convolutional Neural Networks

Side-scan sonar is widely used in underwater rescue and the detection of undersea targets, such as shipwrecks, aircraft crashes, etc. Automatic object classification plays an important role in the rescue process to reduce the workload of staff and subjective errors caused by visual fatigue. However, the application of automatic object classification in side-scan sonar images is still lacking, which is due to a lack of datasets and the small number of image samples containing specific target objects. Secondly, the real data of side-scan sonar images are unbalanced. Therefore, a side-scan sonar image classification method based on synthetic data and transfer learning is proposed in this paper. In this method, optical images are used as inputs and the style transfer network is employed to simulate the side-scan sonar image to generate “simulated side-scan sonar images”; meanwhile, a convolutional neural network pre-trained on ImageNet is introduced for classification. In this paper, we experimentally demonstrate that the maximum accuracy of target classification is up to 97.32% by fine-tuning the pre-trained convolutional neural network using a training set incorporating “simulated side-scan sonar images”. The results show that the classification accuracy can be effectively improved by combining a pre-trained convolutional neural network and “similar side-scan sonar images”.