Recognition of Underwater Materials of Bionic and Natural Fishes Based on Blue-Green Light Reflection

Abstract

Thanks to the advantages of low disturbance, good concealment and high mobility, bionic fishes have been developed by many countries as equipment for underwater observation and data collection. However, differentiating between true and bionic fishes has become a challenging task. Commonly used acoustic and optical technologies have difficulty in differentiating bionic fishes from real ones due to their high similarity in shape, size, and camouflage ability. To solve this problem, this paper proposes a novel idea for bionic fish recognition based on blue-green light reflection, which is a powerful observation technique for underwater object detection. Blue-green light has good penetration under water and thus can be used as a signal carrier to recognize bionic fishes of different surface materials. Three types of surface materials representing bionic fishes, namely titanium alloy, carbon fiber, and nylon, are investigated in this paper. We collected 1620 groups of blue-green light reflection data of these three kinds of materials and for two real fishes. Following this, three machine learning algorithms were utilized for recognition among them. The recognition accuracy can reach up to about 92.22%, which demonstrates the satisfactory performance of our method. To the best of our knowledge, this is the first work to investigate bionic fish recognition from the perspective of surface material difference using blue-green light reflection.

Keywords: bionic fish recognition; blue-green light reflection; machine learning; surface materials; underwater observation.

Figure 1

Schematic diagram of the whole process. The lower left shows the real fishes species and materials representing underwater bionic fishes we selected. On the upper left is the data acquisition device we built. The underwater Y-type optical fiber connects the spectrometer and the tungsten halogen lamp light source. The optical fiber head probing into the water is the transmitting end and receiving end of the light. The sliding platform fixing the optical fiber head is provided with a scale, which can read the light propagation distance. The data set we built is shown on the upper right, with a total of 1620 groups of data recorded to form the data set. The lower right are the machine learning methods we selected to verify the validity of the data set.

Figure 2

The hardware system developed for bionic fish recognition.

Figure 3

We chose nylon, one of the five materials, as a representative to show the difference in distance between the two data forms. (A) Light intensity $I_{ref}$ of nylon collected at different distances. (B) Reflection coefficient $R_C$ of nylon collected at different distances.

Figure 4

Average data ($R_C$) of various materials in clean water and water with salinity of 32‰, the type of material is marked in the subgraph. (A) Larimichthys crocea. (B) Sea bass. (C) Nylon. (D) Titanium alloy. (E) Carbon fiber.

Figure 5

The hyperparameter C value and accuracy curve of SVM and LR, respectively.

Figure 6

The BP neural network we constructed in this paper. The input data contains 222 data points. There are 110 neurons in the first hidden layer and 20 neurons in the second hidden layer. The final output is mapped to five categories.

Figure 7

Confusion matrix of various methods. (A) SVM; (B) LR; (C) BP. The longitudinal axis is the real value and the transverse axis is the predicted value.