| Object detection is one of the hot issues in the field of computer vision,and underwater images are an important means for underwater robots to monitoring ecology and complete underwater operations.Due to the loss of light during underwater propagation,underwater images show serious color shifts and blurred details.In addition,as a result of the influence of water turbidity and different target posture,there are many challenges in the object detection of underwater images.Traditional image object detection algorithms are usually based on region proposal,manual features and feature classification.However,because of the blurred details of underwater images and differences of underwater object postures,the detection effect and realtime of traditional image object detection algorithms is poor.With the development of deep learning,neural networks for object detection based on convolutional neural networks are widely used in various tasks for their abstract feature,robustness,and real-time.This paper mainly studies object detection algorithm for underwater images based on deep learning.The main contents of this paper is as follows:1.This paper studies the basic principles of convolutional neural networks,explains the basic composition of convolutional neural networks,and introduces the activation functions and loss functions commonly used in convolutional neural networks.Meanwhile the receptive field of convolutional neural networks is explained.In addition,this paper also analyzes some typical convolutional neural network structures,and analyze the ideas and significance of their improvements.2.In this paper,the neural network for object detection is divided into two parts.In the part of two-stage neural network,the regional convolutional neural network,which is mainly studied in this paper,is analyzed from the aspects of development process and basic principles.In the part of one-stage neural network,through the development process of the YOLO network,a variety of key technologies used in the YOLOv3 network are pointed out,and the basic principle of the single shot multibox-detector is expounded through the forward propagation process.Through the research on the object detection network algorithm,the basis for the object detection of underwater optical images is provided.3.For the object detection task of underwater image,this paper combines the method of transfer learning to build the neural network for object detection,and improves the regional convolutional neural network mainly studied in this paper.First,in order to adapt to the characteristics of underwater optical image data sets,this paper uses a two-step clustering algorithm to obtain anchor parameters,which increases the reliability of network parameters while ensuring the detection precision.Secondly,in order to improve the detection ability of the network,this paper performs object detection based on the feature pyramid network,and improves the feature pyramid network to enhance the detection efficiency while improving the detection performance.Then,this paper analyzes and improves the fully connected layer of the network,thereby optimizing the total number of the network parameters.Finally,a comparative experiment was conducted between the regional convolutional neural network and typical object detection network,and the detection results and the comparison of the detection precision prove that the detection performance of the improved network in this paper is effective.4.In addition to analyzing network performance through visual results,a reliable object detection network also needs to be evaluated through evaluation indicators.This paper studies the basic evaluation indicators of deep learning,and expounds the calculation methods and meanings of commonly used evaluation indicators in object detection.This paper firstly uses ablation experiments to illustrate the effect and significance of each part of the improved neural network.Secondly,this paper intuitively compares the detection performance of neural network for object detection in the underwater image through the precision-recall curve.Finally,the effectiveness of the improved neural network in this paper is proven through the mean average precision under different conditions. |
