1 Introduction

New discoveries such as mineral resources, energy resources, and marine life in the deep-sea have the great potential to impact our lives in beneficial ways so that there are many good reasons for researchers from various fields to be attracted to such investigations. However, we are still unable to obtain detailed information on the mechanisms of the marine ecological system because the high pressure, darkness, and radio attenuation of the deep sea prevent direct access. Modern computer development, however, makes it possible to use underwater robots to observe the deep-sea [1, 2]. In particular, Autonomous Underwater Vehicles (AUVs) are anticipated as the most important and effective tool for wide area observation since they can cruise for long distances keeping constant depth or altitude with stable posture [3–5]. To achieve a specific mission, AUVs should have high autonomy through various onboard devices such as powerful computers, an accurate sensory system for navigation (INS, DVL), actuators for motion control, and special mission-dependent sensors [6].

As a next-generation AUV, we are developing a Sampling-AUV that can dive into the deep sea to retrieve samples of marine life, aiding in our quest to gain a more comprehensive understanding of marine ecosystems. In the mission scenario, the Sampling-AUV is required to transmit deep-sea floor images to the scientists monitoring from a research ship using acoustic communication. The scientists will then select the marine creatures to sample and the AUV is tasked with retrieving them. The AUV must maneuver to the area where the marine creatures of interest were observed, collect and bring the samples back. The mission scenario, thus, consists of the following steps: (i) deep-sea floor image acquisition, (ii) interesting image selection, (iii) image transmission to research ship, (iv) marine life recognition by scientists, (v) AUV navigation to the target area of interest, (vi) sampling the targets, and (vii) surfacing to bring back samples to the research ship.

For this mission scenario to succeed, the sea-floor images should be enhanced to assist the judgement of the scientists since the color red attenuates rapidly, causing the images to become bluish. Light attenuation also affects image brightness due to small differences in AUV altitude to the sea-floor. And despite slow and inaccurate underwater acoustic communication, the AUV has to select and retrieve targets which include marine life.

Regarding automated marine life recognition, Bewley has proposed an automated species detection method which is based on supervised learning using specific features and the checking of a whole image area to detect coral reefs [7]. Boudhane has proposed an image localization and detection method which includes image enhancement by pre-processing and region segmentation of the images via a mean shift algorithm for fish detection [8]. The pre-processing is based on Poisson–Gauss theory to remove dotted noise such as marine snow to improve fish detection. The challenge of these research is to realize good performance in target detection even when targets change their shapes against their photographed position, pose and individual differences. In these works, image enhancement plays an importance role in improving target recognition.

In this paper, we propose a deep-sea floor image enhancement method based on the Retinex theory, and the performance of image enhancement is evaluated through crab recognition.

2 Problem statement and related works

2.1 Problem statement

In this section, the problems associated with deep-sea floor images taken by AUVs are discussed. An overview of the deep-sea floor image acquisition process and captured deep-sea floor images at different locations and altitudes are shown in Fig. 1.

Fig. 1
figure 1

Overview of deep-sea floor observation by AUV and photographed images in different altitudes: a Changes in AUV altitude due to water steam or rapid change of sea floor altitude, b Ia (altitude: 0.5 m), c Ib (altitude: 1.9 m), d Ic (altitude: 2.2 m), e Ic (altitude: 2.5 m)

Full size image

The AUV altitude \({d}_{i}\)changes due to disturbances such as a water stream or the rapid change of the sea floor, as depicted in Fig. 1a. For example, altitudes \({d}_{1}~{d}_{3}\) at three positions \({p}_{1}~{p}_{3}\) have influence on the color properties of the obtained images due to light attenuation. In general, long wavelength lights such as red and orange are absorbed quickly in water. The intensities of the red, green, and blue lights in pure water after traveling a distance of 10 m are known to be about 1, 67, and 71%, respectively [9].

Images Ia to Id in Fig. 1(b)–(e) were taken at a gas hydrate area near Jouetsu City, Japan using the AUV “Tuna-Sand” [10], and we could observe different color intensities. As the altitude increased, non-uniform illumination of the images increased, as can be seen in Ic and Id, and the color intensities also decreased. If the distance to the sea floor is short as in Ia, a shadow appears in the image even though color saturation is higher than those in the other images. Ib shows uniform illumination and low color saturation compared with Ia. The attenuation effect on each image was analyzed using the HSB (Hue, Saturation, and Brightness) color models, as shown in Fig. 2.

Fig. 2
figure 2

Hue and Saturation (HS) graph and Brightness (B) histogram of corresponding images Ia–Id in Fig. 1 a HS graph of Ia, b HS graph of Ib, c HS graph of Ic, d HS graph of Id, e B histogram of Ia, f B histogram of Ib, g B histogram of Ic, h B histogram of Id

Full size image

It can be seen that color saturation (HS graph) of Ia is distributed in all degrees, and the color of the pixels is located mainly on the color of the crabs (red color area: 330~30[deg.]) and sea floor (cyan and blue color area: 150~270[deg.]), as shown in Fig. 2a. Distribution around the red color area indicates the presence of crabs, providing useful information for creature identification. In the Ia image, the color distribution (hue and saturation) is clear enough to separate the target crabs and sea floor area due to the low influence of light attenuation. Figure 2b, c, d shows the HS graphs of Ib, Ic, and Id, respectively, with the attenuated color values by image altitude. Figure 2b, c shows low color saturation around the red color area, while Fig. 2d does not show any values around the red color area due to light attenuation.

Figure 2(e), (f), (g) and (h) shows the brightness histograms of Ia, Ib, Ic, and Id, respectively. In Fig. 2e, the light source is too close to the sea floor and crabs, and shading appears, contributing to the widely distributed brightness values of Ia. Figure 2f shows the assembled brightness values which are around 0.45, meaning Ib has almost uniform illumination. Figure 2g, f shows non-uniform illumination distributed widely only in the dark areas.

Color attenuation and non-uniform illumination are serious problems which make it difficult for automatic target detection and recognition. Therefore, deep-sea floor images must be enhanced to have natural color distribution and uniform illumination as pre-processing applications in order to accurately recognize targets.

2.2 Related works

Target recognition is a major and important field in Computer Vision, and various feature-based algorithms with high recognition accuracy have been proposed in relevant literature [11]. However, deep-sea floor images taken by AUVs have difficulties caused by light attenuation depending on the wavelength and non-uniform illumination, etc. Therefore, pre-processing of dark and non-uniform illuminated deep-sea floor images is required for the target recognition process.

McGlamery has proposed a theoretical foundation model for an underwater camera system which improves the color imbalance of underwater images resulting from light attenuation [12], and Jaffe has proposed an extended algorithm which takes into account the problem of light scattering [13]. The Jaffe-McGlamery model is an image formation model for underwater images consisting of three reflection components: the direct component, forward-scattering component, and backscattering component. The first, the direct component, is light that is directly reflected on the surface of an object. The second, the forward-scattering component, is light that is reflected by an object and has been scattered at a small angle. The backscattering component is light reflected by objects not on the target scene but that enters the camera. The Jaffe-McGlamery model is effective in enhancing underwater images when the three components can be estimated; however, it is difficult to know the components in real-time and online during deep-sea exploration using the computer system mounted on an AUV.

Yamashita has proposed a color registration method for underwater images to enhance light attenuation [14]. In this method, a set of target images taken from different distances are used to estimate the strength of attenuation for each color, and the images are enhanced to have similar color properties in open air conditions. It showed good performance in achieving images close to the originals. However, in our AUV sampling system, the distances to the target are not accurate and there is no guarantee of obtaining several images of the target as the AUV is always cruising during observation mode. The distance to the sea floor or the altitude of the AUV is 2–3 m; thus, non-uniform illumination is seen in the images.

Bazeille has proposed an underwater image enhancement method to solve previous difficulties and improve image quality by several independent filtering processes [15]. This method uses homomorphic filtering to remove the effects of non-uniform illumination, wavelet transformation to suppress noise, anisotropic filtering to enhance edges, and equalizing RGB channels to suppress and adjust predominant coloring. The Bazeille’s method shows enhanced color and contrast in images; however, the enhanced colors certify not like the colors in air condition. The color is one of the important features to recognize creatures. Therefore, it is difficult to mount on a sampling -AUV system for recognition creatures.

In order to realize autonomous surveys by sampling-AUVs, an image enhancement method to correct problems such as light attenuation, non-uniform illumination, and color reconstruction process is necessary. Here, the Singh method [16] and Retinex theory [17, 18] to solve these problems have been compared. The concept of each method is similar in that the photographed images are composed by illumination and reflectance, and illumination needs to be removed for natural images. In the Retinex theory, illumination is estimated using a Gaussian filter, and in the Singh method, illumination is estimated using the fourth order polynomial approximation. The Singh method assumes that the light pattern is formed by a single strobe and illumination changes gradually, fitting the surface image by a fourth-order polynomial. Thus, rapid intensity variations such as shadows occur and it becomes difficult to employ the fitting function to estimate the illumination in the neighborhood of the shadows and edges in the image. As the Retinex method uses the Gaussian filter, the number of strobes is not limited to one if the appropriate σ of a Gaussian is selected. Therefore, in the proposed method, we have used the Retinex theory to enhance the images, and a detailed comparison of the Singh method and Retinex theory, employing a template color board for the images in a pool environment, is explained in detail below.

3 Enhancement of deep-sea floor image

3.1 Image enhancement method

Pre-processing for image enhancement is composed of (a) Retinex to eliminate the influence of non-uniform illumination, (b) contrast enhancement to improve the image contrast (c) hue adjustment to recover the transformed hue values by light attenuation underwater, and (d) brightness leveling to improve image brightness (Fig. 3).

Fig. 3
figure 3

The proposed image enhancement process

Full size image

The Retinex theory [17] [18], a human-inspired vision system, shows color perceptions under different light colors and intensity levels. It was introduced by Land as an image enhancement method for color constancy. In the theory, an image is composed of illumination and reflectance. The illumination is a component brought by the light source and reflectance is a component which is composed of the target’s actual color. The illumination is estimated using a Gaussian filter which is applied to the raw image in the Retinex theory. A natural image is obtained by eliminating illumination from the raw image and by scaling as the pixel values of the image become smaller after elimination. The mathematical equation of the single scale Land Retinex algorithm is expressed in Eq. [1].

$$R(x,y) = \alpha \cdot \{ \log (I(x,y)) - \log (I(x,y) * F(x,y)\} + \beta ,$$
(1)

where I denotes the input image; x and y are the matrix coordinates; α is the gain factor; β is the offset parameter, and the symbol “*” represents the convolution operator. F is a Gaussian surround function defined by Eq. [2], and R is the Retinex output.

$$F(x,y) = K \cdot \exp [ - (x^{2} + y^{2} )/\sigma ^{2} ],$$
(2)

where σ is the standard deviation of the filter and controls the amount of spatial effect, while K is the normalization factor. The Retinex model is introduced as a solution to enhance deep-sea floor images; however, the automatic enhancement of deep-sea floor images at different altitudes is not easy due to the scaling parameters α and β. In the Retinex theory, these scaling parameters are decided heuristically. The gain parameter (α) controls the image contrast and the offset parameter (β) controls the image brightness. The raw images, which have different properties depending on the altitude where the images were taken such as Ib and Id in Fig. 1 quire different parameters to give similar RGB values. If two images including the same object but having different RGB values is checked, one can still recognize the object in the images since human eyes have the ability to adapt. However, changes in an object’s RGB values are difficult to recognize by computer vision. Therefore, an adaptable scaling method which can replace scaling by α and β is necessary for similar RGB levels in each image. In the propossed enhancement method, α and β are set at 128 to normalize the values which are located between 0 and 255.

The contrast and brightness of R are enhanced by contrast enhancement in Eq. [3] (Fig. 3b) and brightness leveling in Fig. 3d.

$$\hat{R}(x,y) = A \cdot R(x,y)^{\gamma } ,$$
(3)

where γ is the parameter related to the correction curve; A is the constant gain (here, 1.0 is used), and \(\widehat{R}\) gives the resultant image by γ correction. γ used in the proposed method is obtained by experiments carried out in a pool, which is explained in detail below. \(\widehat{R}\) represents a uniform illumination image which has high contrast. \(\widehat{R}\) is then divided into three images, the R-, G-, and B-images, to deal with the differences in frequency-dependent light attenuation. The hue values in each image are rotated to be the same with those in the air, respectively, as shown in Eq. [4].

$$\hat {H}_i = \hat{R}_{i(H)} + \theta_{i},(i \in R - {\text{image}},G - {\text{image}},B - {\text{image}})$$
(4)

\({\widehat{R}}_{i\left(H\right)}\) are the hue planes of \({\widehat{R}}_{i}\); θ i are the parameters that rotate the hue, and θ i are decided by calibration experiments in a pool environment using a template color board (details are shown in Sect. 3.2). The new R-, G-, B-images (\({\widehat{H}}_{i}\)) are merged after hue adjustment, as shown in Eq. [5].

$$E = {\hat H_{{\text{R - image}}}} + {\hat H_{{\text{G - image}}}} + {\hat H_{{\text{B - image}}}}$$
(5)

E is the hue-adjusted image which has high contrast and enhanced hue values. As the final process, brightness leveling is performed to give an even distribution of the brightness values centered at 0.5 (center of normalized brightness value) and the process is shown in Eq. [6]:

$${\hat E_{({\text{B}})}}(x,y) = 0.5{E_{({\text{B}})}}(x,y)/E_{({\text{B}})}^{{\text{Avg}}},(\hat E \in {\hat E_{({\text{H}})}},{\hat E_{({\text{S}})}},{\hat E_{({\text{B}})}})$$
(6)

E (B) is the brightness plane of the resultant image by the hue adjustment process in the HSB color space, and \({E}_{\left(B\right)}^{Avg}\) denotes the average of E (B) . \({\widehat{E}}_{\left(B\right)}\) represents the brightness after leveling, and \(\widehat{E}\) is the resultant image after all processing, as shown in Fig. 3.

3.2 Calibration and evaluation experiment

The calibration and evaluation experiment in the pool in a dark environment was carried out to determine the γ and θ i. We also evaluated the enhanced image color by changing the photographed altitude. Furthermore, estimated illuminations by the Singh method and Retinex theory were evaluated. In this experiment, a CCD camera has a resolution of 1280 × 960 was used, and the light source is a LED light of 500 lm and 40° beam angle. First, the illumination images by the Singh method and Retinex theory were compared, as shown in Fig. 4.

Fig. 4
figure 4

The experimental setup and the resultant images by the Singh method and Retinex theory

Full size image

Figure 4a shows the details of the experimental setup. The camera and LED light position are set at the color board center. The template color board, which is composed of four colors: white, red, green, and blue placed on a black board, was observed at different altitudes d (1~3 m, at 0.5 m intervals) in a dark environment with one light source. Figure 4b shows that the resized raw images (640 × 480), c are the estimated illuminations by the Singh method and d is the estimated illuminations by the Retinex theory. Figure 4e shows the enhanced images using c, and f shows the enhanced images using the proposed method in Sect. 3.1. The σ is 128 and the γ is determined to be 0.25 by this experiment. To determine γ, the saturation values in the white, red, green, and blue square parts of the template color board were evaluated at each altitude, and γ was set at 0.25 to have all saturation values over 0.5.

For details, the differences in illumination by each method are shown in Fig. 5.

Fig. 5
figure 5

Results of illuminations by each method

Full size image

Figure 5 shows a comparison of the luminance intensities and illuminations along the lines of the images (d: 2 m). Figure 5a, b shows that the luminance intensity of the illumination image by the Retinex theory changes more gradually than that by the Singh method and have the convex shape with one peak, however, that by the Singh method has two peaks. We, thus, decided to use the Retinex theory for image enhancement as its properties satisfy our assumptions that the light of the AUV is uniform and illumination changes gradually.

The enhanced images by the proposed method as seen in Fig. 4f show vivid color compared to the raw images. Figure 6 shows the hue values in the red, green, and blue areas on the color board at each altitude.

Fig. 6
figure 6

The variations in hue values by altitude on each color space by the proposed method

Full size image

The angle shows the hue value and radial distance in relation to the altitude in underwater conditions. The dotted lines show the hue value of each color area, red: 0, green 145, and blue: 204, respectively. That is, the dotted line is the value of the red, green, and blue on the color board in open air. “x” represents the hue values in the raw image (Fig. 4b); “∆” represents the hue values without hue adjustment \(\widehat{R}\) (in Eq. [3]), and “o” represents the hue values in the enhanced image (Fig. 4f).

As the underwater images are bluish due to frequency-dependent attenuation, the hue values of the raw image of “red on the color board” change to other colors, whereas those of “green on the color board” and “blue on the color board” take similar values to the altitude change. Moreover, the blue-green colors were predominant in the illumination images as compared to the red color so that they were removed in the enhanced image \(\widehat{R}\). Adjustment of the hue values by dividing an image to R, G, and B-images and giving different hue rotations showed better performance without frequency-dependent adjustment. In this calibration, we obtained the parameters for hue rotation by the differences between the dotted lines and the average of ∆ (θ R:0, θ G:18, θ B:−18[deg]).

The results of contrast enhancement and brightness leveling for the images are shown in Fig. 7.

Fig. 7
figure 7

Variations in the average of the saturation and brightness values by the proposed enhancement method

Full size image

In Fig. 7, the X-axis is the altitude and Y-axis is saturation and brightness. The saturation values are over 0.5 in each color, and especially the red color is improved. Furthermore, the brightness values were improved.

The sample images in Fig. 1, Ib, Ic, and Id were processed through the proposed image enhancement algorithm, and the resultant images are shown in Fig. 8.

Fig. 8
figure 8

Enhanced images by the proposed method corresponding to the images in Fig. 1 a Enhanced image of Ib, b Enhanced image of Ic c Enhanced image of Id

Full size image

Comparing the raw images in Fig. 1, we can find the crabs easily since the enhanced images have uniform illumination, and the color and contrast are also highly enhanced.

The corresponding HSB values are shown in Fig. 9 to analyze enhancement performance.

Fig. 9
figure 9

Hue and saturation graph and brightness histogram of the enhanced images in Fig. 8 a HS graph of enhanced image of Ib, b HS graph of enhanced image of Ic, c HS graph of enhanced image of Id, d B histogram of enhanced image of Ib, e B histogram of enhanced image of Ic, f B histogram of enhanced image of Id

Full size image

The HS graphs in Fig. 9a–c show that the colors of the pixels are widely spread. This means that the colors in the enhanced images have more distinguishable colors compared to the raw images in Fig. 1. The histograms of the brightness levels in Fig. 8d–f show the Gaussian shape around 0.5, i.e., image enhancement and brightness leveling performed as expected.

Whole image processing using Id is described in Fig. 10, and the hue and saturation values are shown in Fig. 11.

Fig. 10
figure 10

The resultant images by each process of the proposed image enhancement

Full size image
Fig. 11
figure 11

Hue and saturation graphs by each process in the proposed image enhancement

Full size image

Figure 10b~e shows the processed image of each process described in Fig. 3. a is input image, and b and c show the processed images by Retinex. Figure 10d shows the processed image without hue adjustment, and e shows the resultant image by the proposed process described in Fig. 3. The d image also shows highly enhanced red values since the Retinex model does not take into account light attenuation, giving rise to colors that are unnatural. The e image has adjusted hue values based on the calibration described in 3.2 and, thus, shows natural colors.

In Fig. 11, the pixels of I show that the values lie between 180 and 240 since I includes the illumination. The pixels of R show the values are concentrated at the center of the graph due to the elimination of illumination. The pixels of \(\widehat{R}\) show a wide range for the saturation values by contrast enhancement. The pixels of \(E\&\widehat{E}\) show the hue values are also adjusted by the calibration experiment depicted in Fig. 6.

Details of the resultant images by brightness histograms are shown in Fig. 12.

Fig. 12
figure 12

The brightness histograms by each process in the proposed image enhancement

Full size image

4 Evaluation of image enhancement by crab recognition

4.1 Components of the recognition system

The recognition algorithm is mainly composed of a segmentation process and classification process. In the segmentation process, selection is performed on the target area separating objects such as crabs from the sea floor in the enhanced image. In the classification process, recognition of the target is performed. The details of each process are described in the following sections.

4.2 Image Segmentation

In image segmentation, the enhanced images are divided into the area of the sea floor and the other target area of the crabs. The flow chart of image segmentation and the results of each process are shown in Fig. 13.

Fig. 13
figure 13

Flow chart of the image segmentation and results of each process a The flow chart of image segmentation, b resultant image by binarization, c resultant image by closing, d resultant image by labeling, e resultant image by checking region size

Full size image

First, the image is binarized by a threshold. The image after enhancement has similar brightness in the entire area so that binarization of the enhanced image shows clearer results than that of the raw image. The result of this process is shown in Fig. 13b. Closing is applied to remove the unnecessary small noise area which contains black dots in Fig. 13b. The closing parameters are determined by approximating the target’s pixel number in the images so that the target area does not disappear by too many closings. The result of closing is shown in Fig. 13c. Labeling is applied to give a number to each black area, and an area color represents the number for each area, as shown in Fig. 13d. The last process in image segmentation is checking the region size (e) to confirm whether it is within the pre-determined region sizes.

4.3 Classification

The Bag of Keypoints (BoK) approach based on vector quantization of the invariant descriptor is used for visual categorization [19]. The BoK algorithm is a well-known classification algorithm for target recognition by the clustering of features. Scale Invariant Feature Transformation (SIFT) descriptors are used to obtain the keypoints and features in the images [20]. The SIFT features are represented as a 128-dimensional vector with topological gradients. The BoK with SIFT is useful when the targets have unique shapes and colors. The candidate regions which have crabs or stones are classified by a BoK algorithm using SIFT features to recognize the targets. Deep-sea floor images taken by AUV include various shapes of crabs with different scales and orientations, and the SIFT feature is suitable for this kind of problem because it is invariant towards changes in orientation and scale. The flow chart for the classifications based on the BoK and its results are shown in Fig. 14.

Fig. 14
figure 14

Flow chart of classifications based on BoK and the processed results a Flow chart of classifications using BoK and SIFT, b Resultant images by classification (Green circle: classified target as a crab, Red circle: classified target as a stone)

Full size image

The process of making a dictionary in Fig. 14a is to generate a database to classify targets off-line using the SIFT features. The k-means is used to classify the features, and each cluster is called visual words in the BoK algorithm. Each visual word in the dictionary is then compared to the obtained features from the input regions of the targets. In Fig. 14b, the green circles are the regions classified as a crab while the red circles are the regions classified as a stone. In this paper, 100 images are used to create the dictionary which has 20 visual words. We evaluated the accuracy of the classifications using Eq. [7]:

$${\text{Accuracy}} = \frac{{TP + TN}}{{TP + FP + TN + FN}}.$$
(7)

TP means True Positive where a crab is classified as a crab. TN means True Negative where a stone is classified as a stone. FP represents False Positive where a stone is classified as a crab. FN represents False Negative where a crab is classified as a stone. For the evaluation, we randomly selected 10 deep-sea floor images taken by the AUV. The accuracy of the recognition algorithm is shown in Table 1. TP, TN, FP, and FN in the first row corresponded to the computed values from Eq. [7]. Results of the evaluation showed that the total accuracy was 84%. The results of crab recognition are shown Fig. 15. The value for each image corresponds to that in Table 1.

Table 1 Accuracy of the recognition algorithm
Full size table
Fig. 15
figure 15

The results of crab recognition

Full size image

5 Discussion

We have introduced image enhancement as a pre-processing method for interesting image selection by an AUV. The proposed image enhancement process shows the enhanced colors against the change in light attenuation by altitude in an underwater environment (Sect. 3.2). In a pool experiment, the hue values showed accumulated values against changes in altitude (Fig. 6), and this property referred to as color constancy is one of the important factors preventing misinterpretation of the targets. The accuracy of crab recognition is compared with and without image enhancement by a crab recognition experiment. Recognition with raw images showed low performance for segmentation and classification due to the attenuated color and non-uniform illumination.

In Fig. 16, a is the raw image, b the saturation plane from (a), and c the image enhanced by the proposed method. Figure 16b shows the separated values between the crabs and sea floor by light attenuation. The pixels of red crabs are highly attenuated in underwater conditions and those of the seafloor have higher values as compared with those of the red crabs. The saturation property of the red crabs was used for segmentation and classification, showing 74% accuracy. However, this property is used only when the targets have colors of high attenuation so that it is difficult to apply into the image of various altitudes. The recognition performance using enhanced image shows an improved accuracy of 84%.

Fig. 16
figure 16

Deep-sea floor images comparing properties a Raw image, b Image of saturation plane in the HSB space, c Enhanced image by Retinex

Full size image

In the some images, the recognition shows low accuracy, especially image number 4 (Fig. 15) which has 50% accuracy. The causes for FP and FN are the lack of keypoints and the problem of occlusion. The lack of keypoints results in misrecognition where regions have fewer numbers of SIFT features and not enough to classify the targets. In order to have better performance, the number of keypoints is important.

6 Conclusions

The development of multi-functional AUVs is one of the most important research topics in the exploration of deep-sea regions. AUVs are also the most practical solution to observe, photograph and conduct sampling in the deep-sea in order to better understand marine life and ecosystems. Next-generation AUV systems are expected to transmit the captured deep-sea floor images using acoustic communication on site and in real time. The transmitted images are required to have as close to natural colors as possible for better recognition. In order to have near-natural colors of the captured images, the deep-sea floor images should be enhanced to solve the effects caused by non-uniform illumination and light attenuation.

The proposed deep-sea floor image enhancement method based on the Retinex theory was applied to the deep-sea floor images taken at different altitudes, and it is shown that the crabs in the images became clearer. These enhanced images were applied to crab identification, where the recognition accuracy was improved to 84% from the 74% obtained by a conventional method. It was concluded that deep-sea-floor images could be accurately enhanced by the proposed method, and the development of such an automatic separation method for the sea-floor and other life forms will greatly contribute to realizing new, high performing Sampling-AUVs.