Research On Underwater Puny Object Detection Algorithm

Underwater object detection technology can be applied to Marine environment monitoring,underwater breeding,underwater robot navigation,Marine exploration,and other fields.However,due to the particularity of the underwater scene,puny objects in water gather and accumulate.They are affected by light absorption and scattering,which easily lead to problems such as fuzzy occlusion and color distortion.As a result,it is still a very challenging problem to develop a high-quality robust underwater object detection algorithm.In addition,because the requirements of underwater application scenarios are not taken into account,the general object detection algorithm is directly adopted,affecting the detection accuracy.In order to solve these problems and improve the accuracy and robustness of underwater object detection,this paper constructs a dataset for underwater object detection.It develops a high-precision robust underwater object detection algorithm based on this data.The main tasks are as follows:(1)In view of the underwater puny objects’ problems of feature coupling,overlap,and accumulation,this paper proposes an underwater object detection algorithm with feature selection based on different spatial features.Firstly,an adaptive weighted fusion pyramid structure is designed.By setting learnable weights,low-resolution details and high-resolution deep semantic information are shared across scales to achieve multi-scale spatial feature selection.Then,with the feature decoupling detection head,the separation of the central feature and the boundary feature are separated.We make the central feature participate in the classification task and make the boundary feature participate in the regression task to achieve feature decoupling.(2)Due to the problem of scarce underwater data and difficult annotation,this paper introduces self-supervised contrast learning.It depends on a visual pretext task and proposes an underwater object detection algorithm based on self-supervised contrast learning.During the process of self-supervised contrast learning,a spatial attention mechanism is set up in the model,which lays a good foundation for model transfer in the underwater object detection task.In addition,via intelligent modeling of channel feature relations,a gated channel normalization module is designed to accelerate the training convergence.Finally,The model transfer is carried out to improve the adaptability of the model for the underwater object detection task.(3)For the problem of low-quality features caused by fuzzy occlusion of underwater puny objects,this paper proposes a multi-task collaborative optimization algorithm for underwater object detection based on self-supervision.Firstly,a local and global fusion backbone network is designed to realize the interaction between local information and global information.It improves the feature extraction ability of the network.Then,the self-supervised method is applied,and the pixel reconstruction branch is added to the contrast learning network which will improve the network’s spatial perception of underwater puny objects.Finally,features in the pyramid are optimized to suppress redundant information and improve detection accuracy.