Self-Supervised Pre-Training Joint Framework: Assisting Lightweight Detection Network for Underwater Object Detection

Abstract

In the computer vision field, underwater object detection has been a challenging task. Due to the attenuation of light in a medium and the scattering of light by suspended particles in water, underwater optical images often face the problems of color distortion and target feature blurring, which greatly affect the detection accuracy of underwater object detection. Although deep learning-based algorithms have achieved state-of-the-art results in the field of object detection, most of them cannot be applied to practice because of the limited computing capacity of a low-power processor embedded in unmanned underwater vehicles. This paper proposes a lightweight underwater object detection network based on the YOLOX model called LUO-YOLOX. A novel weighted ghost-CSPDarknet and simplified PANet were used in LUO-YOLOX to reduce the parameters of the whole model. Moreover, aiming to solve the problems of color distortion and unclear features of targets in underwater images, this paper proposes an efficient self-supervised pre-training joint framework based on underwater auto-encoder transformation (UAET). After the end-to-end pre-training process with the self-supervised pre-training joint framework, the backbone of the object detection network can extract more essential and robust features from degradation images when retrained on underwater datasets. Numerous experiments on the URPC2021 and detecting underwater objects (DUO) datasets verify the performance of our proposed method. Our work can assist unmanned underwater vehicles to perform underwater object detection tasks more accurately.

UTD-YOLO: underwater trash detection model based on improved YOLOv5

“…DCN-bottleneck is an important component of CASDRM, and the backbone network's capacity to extract features is determined by the number of DCN-bottlenecks in the CASDRM. In the backbone network, four CASDRM modules are present, and the numbers (3,6,9,6) denote the number of DCN-bottleneck contained in each CASDRM module in the backbone network in turn, as shown in Table 4.…”

“…Regarding the issue of image quality, Wang and Yu 5 improved the low light and high turbidity of underwater images by self-attention mechanism and multistage fusion. To address the issues of color distortion and unclear target features in underwater pictures, Wang et al 6 suggested an effective self-supervised pretraining joint framework based on the underwater self-encoder transform.…”

Underwater Rescue Target Detection Based on Acoustic Images

“…Ye et al [ 6 ] attempted to classify underwater targets in side-scan sonar images using pre-trained VGG11 and ResNet18 and proposed a training sample pre-processing method with better implications for the migration learning effect. The transformation from low-complexity tracking raymaps to real sonar images is learned by training GAN (Generative Adversarial Networks) to randomly generate sample data with a consistent distribution of the training dataset [ 7 , 8 , 9 , 10 , 11 , 12 , 13 , 14 ]. Zhang Wenwu et al [ 14 ] used a GAN model and a CycleGAN model to produce sonar images directly from noisy data in order to expand the acoustic image dataset, and the generated results were not good due to poor image data.…”

“…Zhang Wenwu et al [ 14 ] used a GAN model and a CycleGAN model to produce sonar images directly from noisy data in order to expand the acoustic image dataset, and the generated results were not good due to poor image data. Huo et al [ 15 ] proposed a semi-synthetic data method mainly for extracting target contours of optical images of ships and aircraft, using a Weibull probability distribution function for the sonar images of Barngrover et al [ 11 ], who preprocessed the optical images and combined them with sonar image features to generate mine-like semisynthetic training data to augment the data set. The complex characteristics of sonar images, such as blurred edges, strong noise, and diverse target shapes, make the data difficult to process, and the above is mainly for objects that are large and distinguishable from their contours and surroundings, such as wrecks, crashed aircraft, and mines.…”

Shape-Guided Detection: A joint network combining object detection and underwater image enhancement together

“…The second type of research approaches, which mostly rely on detection algorithms existing in use, focuses on the deliberate design of modules to address the issue of underwater object detection. For example, Chen [3] used high-resolution and semantically rich feature maps to build a network to significantly improve the accuracy of underwater small object detection; Wang [39] proposed a weighted plus ghost-CSPDarknet and a simplified PANet to improve the underwater object detection performance; Liu [24] proposed domain invariant modules and invariant risk minimization penalties to improve the generalization performance of underwater object detection. Fu [6] searches for transferable prior knowledge from detector-friendly images and uses prior knowledge that can guide the detector to eliminate interfering factors.…”