CNN-LSTM 在日本鲭捕捞渔船行为提取中的应用.

Fishing vessels have been ever-increasing to adopt the monitoring system for the combination of human and electronic observers, particularly with a high level of automation in the world. The data processing has required a simple configuration in the electronic monitoring system at present. It is urgent to improve the automation degree of the electronic monitoring system and the management efficiency of relevant fishing. In this study, the combined Convolutional Neural Network (CNN) with the Long Short Term Memory module (LSTM) was used to apply to the traditional electronic monitoring systems for the data extraction from the fishing vessel behavior of Japanese mackerel fishing in the light fishery. Nine types of fishing operations were divided, including sailing, putting net, waiting, pulling net, waiting for fish, organizing fish box, fish in, fish picking and reprinting, according to the operational characteristics of Japanese mackerel fishing boats. Four groups of parallel experiments were designed to compare the performance of a 3-layer CNN, a 3-layer CNN with an LSTM module (3-layer CNN-LSTM), an 8-layer CNN, and an 8-layer CNN with an LSTM module (8-layer CNN-LSTM) in a Japanese mackerel fishing boat during behavior extraction. Both CNN with LSTM modules were designed before the fully connected layers. The behaviors of nine fishing vessels presented the specific time characteristics, such as “pulling net” and “putting net”. Therefore, the 100 consecutive frames were then stitched horizontally to design the dataset during data processing. The frames per second of the video data was 25, and 100 was determined as the length of each dataset. The reason was that the human eye was normally observed the EM video data for 2-4 s to distinguish the two behaviors. Each data sample also contained certain time dimension information in the LSTM module. The test results show that the comprehensive evaluation index, the F1 score of the models that trained by the 3-layer CNN, 3-layer CNN-LSTM, 8-layer CNN, and 8-layer CNN-LSTM in the test set reached 0.794, 0.799, 0.966, and 0.972, respectively. In addition, the average detection speed of each model on three devices was measured to fully detect the detection efficiency of each model, including a laboratory supercomputer, small mobile workstation, and personal computer. The test samples of 5 000 data sets were taken 34.66, 34.50, 37.41, and 37.61 ms to process each piece using 3-layer CNN, 3-layer CNN-LSTM 8-layer CNN, and 8-layer CNN-LSTM on the supercomputer, respectively. In a small workstation, they were 85.47, 86.10, 120.13, and 121.15 ms in PC2, respectively, whereas, 93.77, 102.27, 118.51, and 124.50 ms in PC3, respectively, in the personal computer. Consequently, the network depth significantly improved the efficiency of the model within the specific range, but the detection time also increased significantly. Specifically, the F1 score of the 8-layer CNN increased from 0.794 to 0.966, while the training time increased from 23.028 to 29.624 h, compared with the 3-layer CNN. The LSTM module enhanced the performance of the network model with the sound detection time. Therefore, the CNN-LSTM modules presented excellent application prospects in the high real-time and high-precision scenarios of electronic monitoring systems. The finding can be expected to improve the automation degree of electronic monitoring systems for the better management efficiency of fishing boats. [ABSTRACT FROM AUTHOR]