TPP-TCCNN-based marine fish identification method

Abstract

The invention provides a TPP-TCCNN-based marine fish identification method, and the method comprises the following steps: collecting an original video containing a plurality of fishes, dividing the original video into a training set and a test set, and carrying out the preprocessing of the fish video information of the training set; performing feature extraction on the preprocessed fish video information to obtain an optical flow image; establishing a dual-channel convolutional neural network, inputting the optical flow image and the RGB image into the dual-channel convolutional neural network with a pyramid pooling layer for training, and obtaining output features, wherein fish information contained in the RGB image is consistent with fish information contained in the original video; performing classification training on the output features through a softmax classification model; and inputting the fish video information in the test set into the final softmax classification model to obtain a fish classification result.

Description

technical field [0001] The invention relates to the technical field of fish identification, in particular to a TPP-TCCNN-based marine fish identification method. Background technique [0002] With the rapid development of computer science and technology, artificial intelligence is applied to various research fields. At present, scientific researchers still need to manually sort the data when they go out to sea to fish and collect sample data. The process is very cumbersome, the scene is more complicated, and the time consumption and cost are very huge. In the field of marine intelligent monitoring, the identification of fish species in the video screen has a Wide application prospects. It can not only develop and utilize fish resources, but also play a very active role in the development of marine fishery production, which has great academic research significance and economic benefits. [0003] In recent years, researchers at home and abroad have conducted a lot of researc...

AI Technical Summary

Problems solved by technology

Among various classification methods, the classification algorithm based on eigenvalues ​​is relatively simple and easy to operate, but the disadvantage is that the recognition accuracy is low; the operation of support vector machine is flexible and changeable, and the classification accuracy is high, but it is difficult to realize large-scale data sets. The training also requires manual selection of feature values; in contrast, the convolutional neural network (Convolutional Neural Network, CNN) has the ability to learn independently, with higher classification accuracy and better robustness.

The disadvantage is that the convolutional neural network needs to obtain a large amount of training data. With the increase of the number of neural network layers and the number of training epochs, the complexity of network calculation will increase, and the training time of the model will increase.

As the network deepens, the accuracy of the training set will decrease, affecting the final classification accuracy

[0004] At present, the basic deep learning network model can only realize the category judgment of videos with a relatively short time-series span, and the problem of low classification recognition rate for videos with a long duration

Images