Sea Cucumber Detection Based On CNN And Its Application In Underwater Vehicle

At present,the monitoring and fishing of sea cucumber cultivation are all completed by divers,which is characterized by high labor intensity,high hidden danger,and serious damage to the health of divers.The underwater vehicle which is safe and reliable has strong adaptability to the environment,and the autonomous monitoring and fishing of sea cucumber by underwater vehicle can promote the development of automation,informatization and intelligence of traditional sea cucumber breeding industry.However,due to the complex and time-varying underwater environment,the detection of sea cucumber has become the primary problem for the automated fishing of sea cucumber.Convolutional neural network has strong ability of feature extraction and expression,and has outstanding performance advantage in target detection under complex background.In this paper,the convolutional neural network is introduced into the detection of underwater sea cucumber,and good detection results are obtained by combining with underwater robot.Firstly,underwater sea cucumber images are collected by underwater vehicle in the marine environment,and the effects of water and illumination on underwater image quality are analyzed.The multi-scale Retinex algorithm is chosen to enhance the images to solve the problem of image attenuation and color cast.Sea cucumber data sets are produced by data augmentation.Secondly,the underwater sea cucumber detection is realiz ed based on the SSD detection algorithm.The K-means clustering is choosed to obtain the a priori width and height of sea cucumber,and the effect of image quality and quantity on the detestion effect is studied.For the network redundancy of the algorithm,the feature extraction network is pruned by filters pruning,and the model size is reduced without loss of precision.Compared with other algorithms,the pruning SSD has obvious performance advantages.To further reduce the size of the model,Mobile Net is introduced into the sea cucumber detection model.For the precision loss caused by lightweight network,a lightweight sea cucumber detection model based on residual network and dilated convolution is proposed to achieve the recognition effect of high pre cision and high speed.Model quantization is applied to transform the data storage format to ompress the model,which lays a foundation for the application of the model.Finally,an underwater vehicle combined with the sea cucumber detection model is designed and installed,and the sea cucumber detection experiments are carried out in the self-built data sets and simulation environment.The experimental results show that the sea cucumber detection model proposed in this paper can identify sea cucumbers efficiently and accurately in complex environments,and has good practicability and reliability.