Marine Fish Recognition Based On Feature Extraction And CNN Model Fusion

The rapid development of China’s fishing industry has led to the phenomenon of overfishing.Meanwhile,a large number of underwater observation data have been produced in the field of ocean observation,which has been booming in recent years.Based on these two points,the demand for automatic identification of Marine fish is increasing year by year,but the unconstrained natural environment makes automatic identification technology face great challenges.Traditional recognition algorithms based on fish image feature extraction have low efficiency and recognition rate.In recent years,with the rapid development of artificial intelligence,deep learning algorithm has brought a new breakthrough to the field of image recognition,but it is also easy to appear the image in the process of network transmission characteristics loss and network over-fitting and other problems.Aiming at the above technical requirements and existing problems,based on the traditional image feature extraction and deep learning algorithm,a high-precision fish recognition framework is proposed.Firstly,fish feature pre-extraction is carried out.Gabor filter and Histogram of Oriented Gradients(HOG)algorithm are used to extract the features of the samples from the original fish image set on which the deep learning depends.The outline and texture features of fish are obtained as part of the training data set.The purpose is to avoid the loss of features in the deep learning by adding important features manually.Secondly,network training is carried out.Affine transformation is used for image data augmentation and size adjustmen",and training set scale is enlarged to improve the training effect.Then the lightweight convolution neural network SqueezeNet is improved to adapt to the training of fish sample set.Training optimization is carried out by network pre-training,batch normalization,adding PReLUs-Softplus(Parametric Rectified Linear Unit-Softplus)function and other methods to deal with network over-fitting,convergence difficulties and other problems.In this way,an ideal network model with high recognition rate can be trained.Finally,in order to further improve the recognition effect,multiple depth neural networks are introduced to train the training set,and different base classifiers are obtained.According to the idea of integration,multiple base classifiers are fused.The final recognition result is the final recognition result of the sample.This makes each image judged by multiple classifiers,and the final recognition result is more accurate and reliable.In the experiment,the image data of marine fishing activity monitoring,Nature Conservancy Fisheries Monitoring,was used as the data set.After introducing the convolution neural network to the training of samples,the marine fish images under eight complex backgrounds are recognized,which achieves a high recognition rate.Firstly,the feature image added enriches the learning range of the neural network,reduces the feature loss of the image,the recognition effect is also better than that of single training original sample.Secondly,the small-scale model brought about by the lightweight network framework based on SqueezeNet also provides great convenience for the embedded deployment of the model in practical applications.Finally,model fusion further improves the recognition rate and stability of the recognition framework,which avoids the poor recognition effect of a single model on a certain kind of image,and makes the recognition rate of the final test sample stable at a higher level.