CN-116310872-B - Luminous remote sensing data light-induced fishing boat detection method based on improved YOLOv network

Abstract

The invention discloses a luminous remote sensing data light-induced fishing boat detection method based on an improved YOLOv network, which uses ship monitoring system data to verify a light-induced fishing boat, marks the light-induced fishing boat on a luminous remote sensing image, constructs a fishing boat target detection data set containing various complex background noises, proposes a YOLOv-based attention mechanism multi-scale fusion algorithm, adopts the algorithm to detect the target of luminous remote sensing image data, comprises a light-induced fishing boat target detection data collection structure, and adopts a YOLOv-based attention mechanism multi-scale fusion light-induced fishing boat detection algorithm, and a light-induced fishing boat position information product production method. The target detection algorithm of the light-induced fishing boat provides technical support for the fishing effort estimation of the ocean and offshore light-induced fishing industry, the space-time variation of the fishing ground, the monitoring of out-of-range fishing application and the like.

Inventors

• CHENG TIANFEI
• ZHANG SHENGMAO
• WU YUMEI
• YANG SHENGLONG
• CUI XUESEN
• ZHOU WEIFENG
• WANG FEI

Assignees

• 中国水产科学研究院东海水产研究所

Dates

Publication Date

20260508

Application Date

20230322

Claims (3)

1. 1. A luminous remote sensing data light-induced fishing boat detection method based on an improved YOLOv network is characterized in that a ship monitoring system (Vessel Monitoring System, VMS) data is used for checking a light-induced fishing boat, the light-induced fishing boat on a luminous remote sensing image is marked, a fishing boat target detection dataset containing various complex background noises is constructed, a attention mechanism multi-scale fusion algorithm based on YOLOv is provided, and target detection is carried out on luminous remote sensing image data by adopting the algorithm, and the method specifically comprises the following three steps: (1) Light induced fishing boat target detection data set structure; (2) YOLOv 5-based attention mechanism multi-scale fusion light-induced fishing boat detection algorithm; (3) A production method of light-induced fishing boat position information products; In order to achieve a better target detection effect, the step (1) needs to construct a plurality of special complex background noise target detection data sets aiming at the characteristics of DNB noctilucent remote sensing images, and the specific operation steps are as follows: (11) Preprocessing the data of DNB noctilucent remote sensing images; (12) Labeling the data set pictures; the step (11) of preprocessing the DNB noctilucent remote sensing image data comprises the following specific operation steps: (111) Firstly, preprocessing all DNB images; (112) Multiplying DNB radiance value by billions and taking logarithms, wherein the step is to enhance the contrast of the features, reduce the complexity of subsequent model calculation and improve the calculation efficiency; (113) Secondly, cleaning and screening VMS information of the light-induced fishing boat to obtain VMS records closest to the shooting time of the DNB sensor; (114) Cutting the 3072×4064 pixel original DNB image to obtain 256×256 pixel slices containing the light induced fishing vessel target, which would be converted into gray scale images for subsequent marking of the vessel's position; the step (113) is to obtain VMS record within 1 hour of the shooting time of DNB sensor, including time, longitude and latitude; the labeling of the data set pictures in the step (12) comprises the following specific operation steps: (121) Labeling the light-induced fishing boat target on the gray level image by using LabelImg tools according to longitude and latitude information provided by VMS information; (122) Comprehensively considering the factors of the ship operation safety distance and the image space resolution, selecting a 5 multiplied by 5 rectangular box for marking, and constructing a data set similar to the PASCAL VOC data set; (123) Each slice of the light-induced fishing boat is provided with a corresponding XML file, and the position of the boat, the type of objects contained in the slice and the size of the boat slice are marked; (124) The data set labeling format can be used for SSD and Faster R-CNN target detection models; (125) The data set labeling format adopted by the YOLO series algorithm is different from the XML format, and the series algorithm generates TXT file after labeling is completed to store id numbers of target categories, x coordinates of a center point of the target/total width of pictures, y coordinates of the center of the target/total height of pictures, width of a target frame/total width of pictures and height of the target frame/total height of pictures; (126) In order to construct a data set which can be applied to an algorithm model of the YOLO series, a conversion process is performed on a markup document of the data set; aiming at the defects that the convolution of YOLOv network structures causes the loss of image key information on a deep feature map and the extraction of the image key feature information is imperfect and insufficient, a attention mechanism module and a weighted bidirectional feature pyramid are added into YOLOv network structures to perform feature fusion, and a detection layer is added at the same time to improve the accuracy of small target detection under a complex background; Aiming at the characteristics of large width of the noctilucent remote sensing image and relatively small light induced fishing boat, the step (3) designs a detection method aiming at the actual noctilucent remote sensing image according to the idea of 'segmentation-detection-combination'; The production method of the light-induced fishing boat position information product in the step (3) comprises the following specific operation steps: (31) Firstly, performing pretreatment of detection model input on DNB and h5 format images; (32) Preprocessing the slices, wherein the preprocessing method is consistent with the constructed detection data set; (33) Multiplying DNB radiance value by billions and taking logarithm; (34) Subsequently converting 180 slices into a gray scale image; (45) Calling a trained target detection model to detect a target; (36) Outputting the detected picture, the normalized center position point coordinates of each detected target and the confidence of the detected target; (37) Converting the coordinate of the normalized center position point of each ship target into longitude and latitude coordinates according to the line number of the slice of the original DNB image and longitude and latitude information; (38) And then removing the wrong targets on land through mask processing to obtain the longitude and latitude data of the ship in the research area.
2. 2. The method for optically-induced fishing boat detection based on noctilucent remote sensing data of the improved YOLOv network according to claim 1, wherein the step (31) is to cut each grid into 256 rows by 256 columns according to a regular grid, and cut the grid into 12 rows by 15 columns for 180 slices in total.
3. 3. The method for optically inducing fishing boat detection based on noctilucent remote sensing data of improved YOLOv network as claimed in claim 1, wherein the gray level image in the step (34) is 8bit deep.

Description

Luminous remote sensing data light-induced fishing boat detection method based on improved YOLOv network Technical Field The invention relates to the technical field of fishery remote sensing information, in particular to a luminous remote sensing data light-induced fishing boat detection method based on an improved YOLOv network. Background The method for carrying out target detection on the photo-induced fishing boat based on noctilucent remote sensing data at home and abroad comprises the steps of convoluting images by wiener filtering and median filtering, removing smiling face noise, salt and pepper noise and the like, adopting a series of automatic methods such as a peak detection algorithm, sharpness index and the like, and detecting formed brightness characteristics of the photo-induced fishing boat on the noctilucent remote sensing image by a threshold segmentation method. The existing target detection method for the light-induced fishing boat based on noctilucent remote sensing data at home and abroad has limitations, and has good detection effect when the lunar illuminance is low. Under full month conditions, a large number of false detections may occur due to the brightness variation of the cloud. The complexity of the environment where the light induced fishing boat is located is mainly represented by weather changes, cloud layers (cyclone, cloud and fog), and the working distance is relatively short, and the light induced fishing boat is gathered together under the influence of the scattering of the cloud layers. A large amount of characteristic information is lost, resulting in a higher omission ratio. The traditional detection algorithm based on the threshold value can not extract effective features under complex conditions, so that the feature expression is imperfect. In addition, the imaging incidence angle, the morning and evening line and the spatial resolution of the sensor are affected, the traditional algorithm is used for extracting the light-induced fishing boat on the noctilucent remote sensing image, and the interference of moon light illuminance and cloud layer optical thickness is large. Disclosure of Invention The invention provides a luminous remote sensing data light-induced fishing boat detection method based on an improved YOLOv network, which aims to accurately extract the light-induced fishing boat from complex background noise such as weather change, intensive operation, cloud layer scattering and the like so as to solve the problem of high-precision small target detection under the complex background noise. The application is realized by the following technical scheme: A luminous remote sensing data light-induced fishing boat detection method based on an improved YOLOv network uses a ship monitoring system (Vessel Monitoring System, VMS) data to check a light-induced fishing boat, marks the light-induced fishing boat on a luminous remote sensing image to construct a fishing boat target detection dataset containing various complex background noises, proposes a attention mechanism multi-scale fusion algorithm based on YOLOv5, and adopts the algorithm to carry out target detection on luminous remote sensing image data, and specifically comprises the following three steps: (1) Light induced fishing boat target detection data set structure; (2) YOLOv 5-based attention mechanism multi-scale fusion light-induced fishing boat detection algorithm; (3) A production method of a light-induced fishing boat position information product. In the step (1), for achieving a better target detection effect, a specific multiple complex background noise target detection data set is constructed according to the features of the Night/Day Band (DNB) noctilucent remote sensing image, and the specific operation steps are as follows: (11) Preprocessing the data of DNB noctilucent remote sensing images; (12) Labeling the data set pictures. As a preferred embodiment, the data preprocessing of the DNB luminous remote sensing image in the step (11) includes the following specific operation steps: (111) Firstly, preprocessing all DNB images; (112) Multiplying DNB radiance value by billion (10-9), and taking logarithm, wherein the step is to enhance the contrast of the feature, reduce the complexity of subsequent model calculation and improve the calculation efficiency; (113) Secondly, cleaning and screening VMS information of the light-induced fishing boat to obtain VMS records (time, longitude and latitude) closest to the shooting time of the DNB sensor (within 1 hour); (114) The original DNB image of 3072×4064 pixels is cropped to obtain 256×256 pixel slices containing the light induced fishing vessel target, which are converted into gray scale images (8 bit depth) for subsequent marking of the vessel's position. As a preferred embodiment, the labeling of the dataset picture in the step (12) comprises the following specific operation steps: (121) Labeling the light-induced fishing boat target on the g