Research On Multi-scale Underwater Object Detection Algorithm

Nowadays,underwater robots are widely used in the tasks of underwater exploration and underwater search.Autonomous underwater robots relies on an underwater object detection algorithm.Underwater object detection is a relatively special application in the field of object detection,it is more challenging due to the complex underwater scenes.Compared with real scenes,underwater scenes are often accompanied by problems such as color domain shift,uneven illumination,blur and distortion.Generic object detection algorithms can hardly maintain high robustness when processing noise-mixed underwater images.To approach the above-mentioned problems,this article conducted a series of researches based on the shallow-sea underwater datasets to explore the object detection algorithm,which adapts to the underwater environment.The main contributions are as follows:(1)Due to the color domain shift and blur in the underwater scene,the pixel-level enhancement of underwater images is carried out by constructing a Multi-image Enhancement algorithm,so that the color domain of the enhanced images are closer to the real-world images,which makes the model improve the generalization ability in the underwater scenario.At the same time,to improves the quality of the proposals of the Region Proposal Network and enhances the accuracy of model recognition,this article propose an Attention Region Proposal Network and introduce the hybrid domain attention mechanism to further enrich the semantic information of the important features.(2)Aiming at the problem that generic single-stage detectors are difficult to detect underwater blurred small targets effectively.This article proposes a Dual refinement underwater object detection network based on the single-stage detection algorithm RFBNet.Firstly,a composite connection feature extraction network with cascaded basic backbones is constructed.Secondly,to gain multi-scale features,the receptive field enhancement module with dilated convolution is introduced,which can achieve the effects of getting a larger receptive field under the same size convolution kernel by controlling the dilation rate of the dilated convolution.Finally,to approach the sample imbalance problems of the single-stage detectors,the Prediction Refinement Scheme is proposed,which introduces deformable convolution to perform two regressions to refine the anchors and features,so that the two can be effectively aligned.(3)To solve the problem of redundant anchors in generic object detectors.Based on Sparse-RCNN,this article constructs a new paradigm without convolutional underwater extraction and applies the Transformer to the underwater scenes.The multi-head self-attention mechanism encoder is used to extract features from the input images,which abandon the traditional feature extraction methods based on convolutional neural networks and use the self-attention layer to make up the feature extraction network and detection heads.To accelerate the convergence of the model,a dynamic instance interaction module is introduced.Through the one-to-one interaction between anchors and proposal features,the candidates are screened,which makes the model optimized in the training phase.