WO2023116631A1 - 旋转船只目标检测模型的训练方法、训练装置和存储介质 - Google Patents

旋转船只目标检测模型的训练方法、训练装置和存储介质 Download PDF

Info

Publication number
WO2023116631A1
WO2023116631A1 PCT/CN2022/140069 CN2022140069W WO2023116631A1 WO 2023116631 A1 WO2023116631 A1 WO 2023116631A1 CN 2022140069 W CN2022140069 W CN 2022140069W WO 2023116631 A1 WO2023116631 A1 WO 2023116631A1
Authority
WO
WIPO (PCT)
Prior art keywords
training
ship
target detection
value
detection model
Prior art date
Application number
PCT/CN2022/140069
Other languages
English (en)
French (fr)
Inventor
周翊民
吴相栋
Original Assignee
中国科学院深圳先进技术研究院
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 中国科学院深圳先进技术研究院 filed Critical 中国科学院深圳先进技术研究院
Publication of WO2023116631A1 publication Critical patent/WO2023116631A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Definitions

  • the invention belongs to the technical field of computer vision, and in particular relates to a training method, a training device, a computer-readable storage medium and computer equipment for a rotating ship target detection model.
  • Unmanned Aerial Vehicle has many advantages such as simple structure, easy maintenance and portability, high efficiency and high reliability of task execution, it can take off and land anytime and anywhere, has strong operational flexibility, and can meet the needs of various emergencies. It is suitable for flexible inspection of ports, fishing grounds, tourist areas, etc.
  • the optical image from the UAV has more detailed information and more obvious geometric structure, which is more intuitive and easy to understand.
  • drones are less disturbed by clouds and fog from the perspective of drones, and clearer target imaging can extract more effective features, and the drone inspection method is more suitable for flexible inspection tasks in specific areas.
  • the ship target on the sea surface has the characteristics of indeterminate direction and large aspect ratio, and it is impossible to achieve high-precision and fast recognition by using general methods.
  • the task of rotating target detection it is mainly divided into two categories, namely the five-parameter method (the method based on the angle) and the eight-parameter method (the method based on the corner point).
  • the main research direction is considering the improvement of anchor box and bounding box generation, feature alignment, extraction of rotation invariant features, etc.
  • directed bounding box generation methods have emerged that use generative probabilistic models to extract OBB proposals. For each proposed region, its location, size and orientation are determined by searching for a local maximum likelihood.
  • the task of ship detection in aerial images faces the problem of data imbalance. There are few data sets and angle information is involved, so it is difficult to provide sufficient sample data.
  • the general data enhancement method has a general effect on this special scene, and the optimization effect on the scale direction is better, but it is difficult to meet the demand for data enhancement of angle information in this scene.
  • the technical problem solved by the invention is: how to overcome the problem of data imbalance in the training process of the rotating ship target detection model.
  • a training method for a rotating ship target detection model comprising:
  • the method for setting the rotation trigger probability of the current training round according to the rotation loss after the last round of training includes:
  • the rotation trigger probability of the current training wheel is set to a first predetermined value; if there is no spin loss imbalance, the rotation trigger probability of the current training wheel is set to a second predetermined value.
  • the enhanced image of the ship is input into the rotating ship target detection model to be trained, and the method for obtaining the predicted value includes:
  • the rotation feature alignment operation is performed on the convolution feature map of each level, and the method of obtaining the convolution area after the feature alignment includes:
  • the convolution area after feature alignment is calculated according to the angle corresponding to the maximum response value.
  • the real value is the center coordinate, width, height and angle corresponding to the real frame of the ship label image
  • the method for calculating the first loss value based on angular distance includes:
  • a weight parameter based on angular distance is calculated according to the angle of the real frame, the angle of the predicted frame, and the aspect ratio of the real frame;
  • the first loss value is obtained by calculating according to the angle-distance-based weight parameter and the intersection-union ratio.
  • the method for calculating the second loss value based on angle classification includes:
  • a second loss value is obtained by calculating according to the angle of the real frame and the angle of the predicted frame.
  • the application also discloses a training device for a rotating ship target detection model, the training device comprising:
  • a preprocessing unit is used to perform preprocessing and angle annotation processing on the original rotating ship image to obtain an annotated image of the ship;
  • the data enhancement unit is used to set the rotation trigger probability of the current training wheel according to the rotation loss after the last round of training, and perform data enhancement processing according to the marked image of the ship to obtain the enhanced image of the ship;
  • a data input unit configured to input the enhanced image of the ship into the rotating ship target detection model to be trained to obtain a predicted value
  • a loss calculation unit configured to calculate the loss function value of the current training wheel according to the predicted value and the actual value corresponding to the ship label image
  • a parameter updating unit configured to update the model parameters of the rotating ship target detection model to be trained according to the loss function value to complete the current round of training.
  • the present application also discloses a computer-readable storage medium, the computer-readable storage medium stores a training program of a rotating ship target detection model, and when the training program of the rotating ship target detection model is executed by a processor, the above-mentioned rotation is realized.
  • a training method for a ship object detection model is also disclosed.
  • the present application also discloses a computer device, which includes a computer-readable storage medium, a processor, and a training program for a rotating ship target detection model stored in the computer-readable storage medium, and the rotating ship target detection
  • a computer device which includes a computer-readable storage medium, a processor, and a training program for a rotating ship target detection model stored in the computer-readable storage medium, and the rotating ship target detection
  • the training program of the model is executed by the processor, the above-mentioned training method of the rotating ship target detection model is realized.
  • the invention discloses a training method and training device for a rotating ship target detection model. Compared with the existing method, it has the following technical effects:
  • This method adopts a new data enhancement strategy, and performs dynamic data enhancement during the training process through loss feedback, which improves the problem of sample imbalance, makes data enhancement more targeted, and improves the efficiency of data enhancement.
  • this method designs a module based on angle channel switching and angle distance, which realizes the accurate extraction of angle-related image features, and reduces the classification loss through the learning of angle distance.
  • the discretized angle classification method of angular distance converts the angle range of 180 degrees into 180 classification dimensions for processing, so that the regression problem of angle information is transformed into a classification problem, and boundary discontinuity can be successfully avoided.
  • the pixel-level IOU is calculated as a loss function, so that the influence factors of IOU are considered in the process of loss calculation, so as to obtain a more accurate matching scoring mechanism, which optimizes the training process, and then considers the rotation angle difference as a weight optimization loss function. Improved detection accuracy for ship targets.
  • Fig. 1 is the overall flowchart of the training method of the rotating ship object detection model of embodiment one of the present invention
  • FIG. 2 is an overall structural diagram of a rotating ship target detection model according to Embodiment 1 of the present invention.
  • Fig. 3 is the functional block diagram of the training device of the rotating ship target detection model of the third embodiment of the present invention.
  • FIG. 4 is a schematic diagram of computer equipment according to Embodiment 4 of the present invention.
  • the rotation trigger probability of the current round is calculated according to the rotation loss after the previous round of training, that is, by introducing a loss feedback mechanism , dynamically adjust the rotation trigger probability in the data enhancement of each round of training process, and perform more targeted data enhancement on the angle information, improve the efficiency of data enhancement, improve the problem of sample imbalance, and then enhance the obtained ship
  • the image is input as a training sample into the rotating ship target detection model to be trained for model training to improve the detection efficiency of the model.
  • the training method of the rotating ship target detection model of the first embodiment includes the following steps:
  • Step S10 preprocessing and angle labeling processing are performed on the original rotating ship image to obtain a ship labeling image.
  • Step S20 Set the rotation trigger probability of the current training wheel according to the rotation loss after the last round of training, and perform data enhancement processing according to the marked image of the ship to obtain the enhanced image of the ship;
  • Step S30 input the enhanced ship image into the rotating ship target detection model to be trained to obtain the predicted value
  • Step S40 calculating the loss function value of the current training wheel according to the predicted value and the real value corresponding to the ship label image;
  • Step S50 update the model parameters of the rotating ship target detection model to be trained according to the loss function value, and complete the current round of training.
  • the ship image data of the relevant scene is obtained and the angle annotation is performed. Afterwards, the image data of the ship is marked according to the angle coordinates and the target inclination angle.
  • the target coordinates and angle expressions used are (x, y, w, h, ⁇ ), where ⁇ adopts the long side definition method, that is, the angle passed by the first long side from the counterclockwise rotation from the horizontal direction, The value range is 0-180°.
  • the data set used for ship image data is HRSC2016, which is a ship detection data set with a large aspect ratio and a wide range of arbitrary directions. This dataset contains two scenes (ships at sea and ships offshore) with 15 object categories. The size of each image ranges from 300 ⁇ 300 to 1500 ⁇ 900.
  • the dataset has 1061 images, of which 436 are used for training, 181 for validation and 444 for testing.
  • step S20 the ship label images to be used in each round of training are enhanced.
  • the way of enhancement is to splicing the four images by random zooming, random cropping, random rotation and random arrangement to form a A new image, an enhanced image of the ship.
  • the enhancement effect is controlled by setting a trigger probability for each enhancement method.
  • This embodiment introduces a loss feedback mechanism. During the training process, various losses are calculated after each training epoch, and the contribution value of each loss is judged.
  • the angle loss contributes less or more (by setting the threshold to judge), that is, it is judged that the rotation loss is unbalanced
  • adjust the trigger probability of rotation enhancement in the above data enhancement process that is, set the rotation trigger probability is the first predetermined value, which alleviates the problem of data imbalance to a certain extent.
  • the data enhancement is still performed according to the originally set trigger probabilities, that is, the rotation trigger probability is set to a second predetermined value. Exemplarily, if it is too low, the rotation trigger probability is increased by 50%, and if it is too high, it is decreased by 50%. This adjustment ratio can be set according to actual needs.
  • the rotating ship target detection model to be trained is based on YOLOv5, and the CSP structure and Focus structure are introduced on the basis of Darknet53 to form the basic Backbone.
  • the spatial pyramid pooling layer (Spatial Pyramid Pooling layer, SPP) and the structure of FPN+PAN are introduced.
  • SPP Spatial Pyramid Pooling layer
  • FPN is a top-down structure, which transfers and fuses high-level feature information through upsampling to obtain a feature map for prediction.
  • a bottom-up feature pyramid module containing two PAN structures is added to FPN, which is able to convey position features.
  • the combination of FPN and feature pyramid structure including PAN performs parameter aggregation for different detection layers from different backbone layers, which further improves the feature extraction ability of the model.
  • the rotating ship target detection model also has a rotating feature alignment module.
  • the rotating feature alignment module uses the maximum pooling method to evaluate the response strength between each feature area and the real area of the target by switching the angle channel, and then obtains the angle value through angle interpolation. The feature alignment effect can make the subsequent extracted features better match the rotating ship target.
  • the convolution operation is performed to form a multi-layer convolution feature map, and the convolution feature maps of different levels are rotated
  • the operation of feature alignment For the pixels in the image, the conventional feature extraction area can be determined through preset anchor frame matching. This embodiment is improved on this basis. For a specific pixel point in a specific anchor frame matching state, the angle channel value is switched, and the corresponding feature map area is calculated for each angle channel value, and the corresponding feature map area and the target true map area are compared.
  • Values corresponding to the feature map area are compared to obtain several direction values whose response is greater than the threshold, and the direction channel value with the strongest response is used as the corresponding angle of feature alignment.
  • the deviation weight between the target true value and the predicted result on the feature map is calculated, and the deviation weight is integrated into the convolution calculation for feature extraction.
  • p is the coordinate position of the operation
  • w and h are the width and height of the rotation box respectively
  • r is the radius of the convolution kernel
  • R( ⁇ ) is the angle rotation matrix
  • W(r) represents the convolution operation
  • X(L) represents the feature map area where the convolution operation is performed
  • Y(p) represents the extracted features.
  • the convolution area L after the feature alignment is regarded as the anchor box, and the offset between the two boxes can be calculated by combining the area covered by the ground truth box and the area covered by the current anchor box, and calculated by their coordinates To get the coordinates of the current prediction frame, the formula is as follows:
  • t x (xx a )/w a
  • t y (yy a )/h a
  • t' x (x'-x a )/w a
  • t' y (y'-y a )/h a
  • x, y, w, h, ⁇ represent the center coordinates, width, height and angle of the truth box respectively;
  • x', y', w', h', ⁇ ' represent the center coordinates, width, Height and angle;
  • x a , y a , w a , h a , ⁇ a represent the center coordinates, width, height and angle of the anchor box respectively, and the symbols t and t' represent the offset.
  • FL means Focal loss
  • L CSL means the second loss value brought by the angle classification method itself.
  • intersection-over-union ratio (IOU, Intersection-over-Union) is adjusted according to the calculated weight parameter of the angular distance to obtain the first loss value, and the specific calculation process is as follows.
  • the first embodiment improves the PIOU method to design a loss function for the rotating object detection task.
  • the calculation of IOU is to evaluate the degree of overlap of two borders. Since the rotation bounding box (Oriented bounding box, OBB) and the intersection area are composed of pixels in the image space, their area is approximately the number of internal pixels. Since OBBs and intersection regions are constructed from pixels in image space, their areas are approximated by the number of interior pixels.
  • the truth box and the prediction box are represented by b and b' respectively.
  • b is the rotation bounding box determined by the values of x, y, w, h, and ⁇ in the above formula.
  • b' is the above The bounding box of the rotation determined by the x', y', w', h', ⁇ ' values in the formula.
  • the binary constraints are defined as follows:
  • d ij is the L 2 norm distance between point p i,j and the OBB center coordinates (x, y)
  • w, h are the width and height of the bounding box
  • d w and d h are the distances along the horizontal direction and The distance d in the vertical direction.
  • intersection-over-union ratio (IOU) of a pair of bounding boxes (b,b') is expressed as:
  • b, b' is regarded as a pair of bounding boxes of a positive sample (the match of this pair of bounding boxes is regarded as a positive sample).
  • PIOU can well represent the degree of overlap of two OBBs, but cannot calculate their angular difference.
  • integrating the angle information into the PIOU method can guide the regression direction more accurately. Construct a weight parameter sensitive to angular distance, and introduce the angle information of the rotating object in the calculation to improve the PIOU loss.
  • the weight parameter can be expressed as:
  • R is the aspect ratio of the real box
  • ⁇ b and ⁇ b' are the real angle and the predicted angle, respectively.
  • the rotation-aware PIoU loss that is, the first loss value based on angular distance
  • M represents all pairs of bounding boxes that are positive samples.
  • This step is to use the cross-entropy loss to adjust the learning of category features in the model through simple feature map matching, and obtain the category-based third loss value L cls ( p n ,t n ).
  • the loss function value formed according to the first loss value, the second loss value and the third loss value is as follows:
  • N in the above formula represents the number of anchor boxes
  • b, b' are the truth box and prediction box respectively.
  • ⁇ n , ⁇ ' n denote the true angle and predicted angle, respectively.
  • t n is the label of the object, and p n is various probability distributions calculated by the Sigmoid function.
  • the hyperparameters ⁇ 1 , ⁇ 2 , ⁇ 3 control the trade-off and default to ⁇ 1,0.5,1 ⁇ .
  • the obtained vessel enhanced image data is divided into training set and test set according to the ratio of about 9:1, and the training set data and initial parameters are input into the target detection model constructed above for forward calculation.
  • the present invention selects the one-stage detection framework YOLOv5l as a benchmark, and adds the above-mentioned dynamic data enhancement method, angle classification idea, feature alignment and other methods into the training process. Comparing the model involved in the present invention with known advanced models such as RoI Transformer and R2CNN, the method designed in the present invention has outstanding overall performance while considering both detection accuracy and detection speed, and it avoids angle regression Periodic problems, and has high real-time and high detection accuracy at the same time, suitable for application in inspection scenarios with high real-time performance.
  • Embodiment 2 also discloses a training device for a rotating ship target detection model, the training device includes a preprocessing unit 100, a data enhancement unit 200, a data input unit 300, a loss calculation unit 400 and a parameter update unit 500.
  • the preprocessing unit 100 is used to perform preprocessing and angle labeling processing on the original rotating ship image to obtain the ship marked image;
  • the data enhancement unit 200 is used to set the rotation trigger probability of the current training wheel, and perform data enhancement processing according to the ship marked image to obtain Enhanced ship image;
  • the data input unit 300 is used to input the enhanced image of the ship into the rotating ship target detection model to be trained to obtain a predicted value;
  • the loss calculation unit 400 is used to calculate the current training value according to the predicted value and the real value corresponding to the marked image of the ship The loss function value of the round;
  • the parameter update unit 500 is used to update the model parameters of the rotating ship target detection model to be trained according to the loss function value to complete the current round of training.
  • the data enhancement unit 200 is also used to determine whether there is a rotation loss imbalance after the last round of training is completed; if there is a rotation loss imbalance, the rotation trigger probability of the current training wheel is set as the first predetermined value; if not If there is a spin loss imbalance, then the spin trigger probability of the current training wheel is set to a second predetermined value.
  • the loss calculation unit 400 is also used to calculate the first loss value based on angular distance based on the center coordinate, width, height, and angle of the real frame and the center coordinate, width, height, and angle of the predicted frame, and classify based on angle
  • the second loss value of and the third loss value based on category, the first loss value, the second loss value and the third loss value constitute the loss function value.
  • Embodiment 3 also discloses a computer-readable storage medium.
  • the computer-readable storage medium stores a training program for the rotating ship target detection model.
  • the training program for the rotating ship target detection model is executed by a processor, the above-mentioned rotating ship target detection is realized.
  • the training method of the model is realized.
  • Embodiment 4 also discloses a computer device.
  • the computer device includes a processor 12 , an internal bus 13 , a network interface 14 , and a computer-readable storage medium 11 .
  • the processor 12 reads the corresponding computer program from the computer-readable storage medium and executes it, forming a request processing device on a logical level.
  • one or more embodiments of this specification do not exclude other implementations, such as logic devices or a combination of software and hardware, etc., that is to say, the execution subject of the following processing flow is not limited to each A logic unit, which can also be a hardware or logic device.
  • a training program for the rotating ship target detection model is stored on the computer-readable storage medium 11. When the training program for the rotating ship target detection model is executed by the processor, the above-mentioned training method for the rotating ship target detection model is implemented.
  • Computer-readable storage media includes both volatile and non-permanent, removable and non-removable media by any method or technology for storage of information.
  • Information may be computer readable instructions, data structures, modules of a program, or other data.
  • Examples of computer readable storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage , magnetic cassettes, disk storage, quantum memory, graphene-based storage media or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by computing devices.
  • PRAM phase change memory
  • SRAM static random access memory
  • DRAM dynamic random access memory
  • RAM random access memory
  • ROM read-only memory
  • EEPROM electrically erasable

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Image Analysis (AREA)

Abstract

本发明公开了一种旋转船只目标检测模型的训练方法、训练装置和存储介质。该训练方法包括：对原始旋转船只图像进行预处理和角度标注处理，获得船只标注图像；根据上一轮训练之后的旋转损失情况设定当前训练轮的旋转触发概率，按照所述船只标注图像进行数据增强处理，获得船只增强图像；将所述船只增强图像输入到待训练的旋转船只目标检测模型中，得到预测值；根据所述预测值和所述船只标注图像对应的真实值计算当前训练轮的损失函数值；根据所述损失函数值更新所述待训练的旋转船只目标检测模型的模型参数，完成当前轮训练。通过loss反馈在训练过程中进行动态数据增强，改善了样本不均衡问题，使得数据增强有了针对性，提高了数据增强的效率。

Description

旋转船只目标检测模型的训练方法、训练装置和存储介质 技术领域
本发明属于计算机视觉技术领域，具体地讲，涉及一种旋转船只目标检测模型的训练方法、训练装置、计算机可读存储介质和计算机设备。
背景技术
无人机(UAV)具有结构简单，便于维护且携带方便，执行任务效率高与可靠性强等诸多优点，其能够随时随地起降，作业灵活性强，能够满足各种突发任务的需求，适合对港口、渔场、旅游区等处的灵活巡检。相对于以往进行舰船识别使用的合成孔径雷达图像，来自无人机的光学图像具有更详细的信息和更明显的几何结构，更直观且容易理解。相较于卫星图片，无人机视角下受到的云雾干扰更少，目标成像更清晰能够提取更多的有效特征，且无人机巡检方式更适用于特定区域的灵活巡检任务。但同时，在无人机视角中，海面船只目标存在的方向不定和大纵横比特征，使用通用方法无法实现高精度快速识别。
目前针对旋转目标检测任务，主要分为两大类，分别为五参数法(基于角度的方法)和八参数法(基于角点的方法)。主要研究方向即考虑锚框和边界框生成改进、特征对齐、旋转不变特征的提取等。在有向边界框生成方面，催生出使用生成概率模型来提取OBB建议的方法，对于每个建议区域，通过搜索局部最大似然来确定其位置、大小和方向。还有研究者设计快速旋转包围框估计算法，实现从任意分割数据集生成旋转包围盒地面真值，利用椭圆拟合来估计带有分割(掩模)的包围盒旋转角度和大小，以训练旋转角度回归模型。此外，特征匹配是目标检测的重点任务，设计带正则化的优化函数来控制旋转映射前后的特征表示是常用方式。针对旋转目标检测领域，最近普遍的方式是在卷积神经网络架构的基础上构建单独的旋转不变层，或者针对航空图像中目标的旋转特性，设计了一个对旋转感兴趣区的学习器，替换传统的水平感兴趣区，从RoI中提取旋转不变特征，这些方式通过对角度通道的处理提高网络结构对目标旋转特征的适应能力，并且使用单独的模块设计，从而能够方便地集成到传统检测器中以提升对旋转目标的检测性能。
当前航空图像中船只检测任务面临数据不均衡问题，数据集较少，且涉及到角度信息，很难提供充足的样本数据。通用的数据增强方法针对该特殊场景 效果一般，主要针对尺度方向优化效果较好，却难以满足该场景下对角度信息进行数据增强的需求。
发明内容
(一)本发明所要解决的技术问题
本发明解决的技术问题是：如何克服旋转船只目标检测模型模型训练过程中数据不均衡的问题。
(二)本发明所采用的技术方案
一种旋转船只目标检测模型的训练方法，所述训练方法包括：
对原始旋转船只图像进行预处理和角度标注处理，获得船只标注图像；
根据上一轮训练之后的旋转损失情况设定当前训练轮的旋转触发概率，按照所述船只标注图像进行数据增强处理，获得船只增强图像；
将所述船只增强图像输入到待训练的旋转船只目标检测模型中，得到预测值；
根据所述预测值和所述船只标注图像对应的真实值计算当前训练轮的损失函数值；
根据所述损失函数值更新所述待训练的旋转船只目标检测模型的模型参数，完成当前轮训练。
优选地，根据上一轮训练的后的旋转损失情况设定当前训练轮的旋转触发概率的方法包括：
在上一轮训练完成之后，判断是否存在旋转损失失衡；
若存在旋转损失失衡，则将当前训练轮的旋转触发概率设定为第一预定值；若不存在旋转损失失衡，则将当前训练轮的旋转触发概率设定为第二预定值。
优选地，将所述船只增强图像输入到待训练的旋转船只目标检测模型中，得到预测值的方法包括：
所述船只增强图像输入到待训练的旋转船只目标检测模型后，得到不同层级的卷积特征图；
对每个层级的卷积特征图进行旋转特征对齐操作，获得特征对齐后的卷积区域；
根据所述特征对齐后的卷积区域确定预测框的中心坐标、宽度、高度和角度，作为预测值。
优选地，对每个层级的卷积特征图进行旋转特征对齐操作，获得特征对齐后的卷积区域的方法包括：
遍历多个角度，计算每个角度通道下，预设锚框在所述卷积特征图中确定的特征图区域与目标真值对应的特征图区域的响应值；
根据最大的响应值所对应的角度计算得到特征对齐后的卷积区域。
优选地，所述真实值为所述船只标注图像的真实框所对应的中心坐标、宽度、高度和角度，根据所述预测值和所述船只标注图像对应的真实值计算当前训练轮的损失函数值的方法包括：
根据所述真实框所对应的中心坐标、宽度、高度、角度和所述预测框的中心坐标、宽度、高度、角度计算得到基于角度距离的第一损失值、基于角度分类的第二损失值和基于类别的第三损失值，所述第一损失值、所述第二损失值和所述第三损失值构成所述损失函数值。
优选地，计算基于角度距离的第一损失值的方法包括：
根据所述真实框所对应的中心坐标、宽度、高度、角度和所述预测框的中心坐标、宽度、高度、角度计算得到所述真实框与所述预测框的交并比值；
根据所述真实框的角度、所述预测框的角度、根据所述真实框的长宽比计算得到基于角度距离的权重参数；
根据所述基于角度距离的权重参数和所述交并比值计算得到所述第一损失值。
优选地，所述计算得到基于角度分类的第二损失值的方法包括：
根据所述根据所述真实框的角度、所述预测框的角度计算得到第二损失值。
本申请还公开了一种旋转船只目标检测模型的训练装置，所述训练装置包括：
预处理单元，用于对原始旋转船只图像进行预处理和角度标注处理，获得 船只标注图像；
数据增强单元，用于根据上一轮训练之后的旋转损失情况设定当前训练轮的旋转触发概率，按照所述船只标注图像进行数据增强处理，获得船只增强图像；
数据输入单元，用于将所述船只增强图像输入到待训练的旋转船只目标检测模型中，得到预测值；
损失计算单元，用于根据所述预测值和所述船只标注图像对应的真实值计算当前训练轮的损失函数值；
参数更新单元，用于根据所述损失函数值更新所述待训练的旋转船只目标检测模型的模型参数，完成当前轮训练。
本申请还公开了一种计算机可读存储介质，所述计算机可读存储介质存储有旋转船只目标检测模型的训练程序，所述旋转船只目标检测模型的训练程序被处理器执行时实现上述的旋转船只目标检测模型的训练方法。
本申请还公开了一种计算机设备，所述计算机设备包括计算机可读存储介质、处理器和存储在所述计算机可读存储介质中的旋转船只目标检测模型的训练程序，所述旋转船只目标检测模型的训练程序被处理器执行时实现上述的旋转船只目标检测模型的训练方法。
(三)有益效果
本发明公开了一种旋转船只目标检测模型的训练方法和训练装置，相对于现有方法，具有如下技术效果：
本方法采用了一种新的数据增强策略，通过loss反馈在训练过程中进行动态数据增强，改善了样本不均衡问题，使得数据增强有了针对性，提高了数据增强的效率。
同时，本方法针对一阶段目标检测的特征对齐问题，设计了一个基于角度通道切换以及角度距离的模块，实现了对角度相关的图像特征的准确提取，并且通过角度距离的学习降低了分类损失。
另外，角度距离的离散化角度分类方法，将180度的角度范围转换为180个分类维度进行处理，使角度信息的回归问题转化成了分类问题，则可以成功避免边界不连续性。同时计算像素级别的IOU作为损失函数，使得损失计算的 过程中考虑IOU的影响因素，从而得到更精确的匹配评分机制，以此优化了训练过程，再考虑旋转角度差，作为权重优化损失函数，提高了对船只目标的检测精度。
附图说明
图1为本发明的实施例一的旋转船只目标检测模型的训练方法的整体流程图；
图2为本发明的实施例一的旋转船只目标检测模型的整体结构图；
图3为本发明的实施例三的旋转船只目标检测模型的训练装置的原理框图；
图4为本发明的实施例四的计算机设备示意图。
具体实施方式
为了使本发明的目的、技术方案及优点更加清楚明白，以下结合附图及实施例，对本发明进一步详细说明。应当理解，此处所描述的具体实施例仅仅用以解释本发明，并不用于限定本发明。
在详细描述本申请的各个实施例之前，首先简单描述本申请的发明构思：现有技术中，航空图像中船只检测任务面临数据不均衡问题，数据集较少，且涉及到角度信息，很难提供充足的样本数据，且通用的数据增强方法难以满足该场景下角度信息的数据增强需求。为此，本申请提供了一种旋转船只目标检测模型的训练方法，在对原始图像数据标注之后，根据上一轮训练之后的旋转损失情况计算当前轮的旋转触发概率，即通过引入损失反馈机制，动态地调整每轮训练过程的数据增强中的旋转触发概率，更有针对性地进行角度信息方面的数据增强，提高了数据增强效率，改善了样本不平衡的问题，接着将得到的船只增强图像作为训练样本输入到的待训练的旋转船只目标检测模型中，进行模型训练，提高模型的检测效率。
具体地，如图1所示，本实施例一的旋转船只目标检测模型的训练方法包括如下步骤：
步骤S10、对原始旋转船只图像进行预处理和角度标注处理，获得船只标注图像。
步骤S20、根据上一轮训练之后的旋转损失情况设定当前训练轮的旋转触发概率，按照船只标注图像进行数据增强处理，获得船只增强图像；
步骤S30、将船只增强图像输入到待训练的旋转船只目标检测模型中，得到预测值；
步骤S40、根据预测值和船只标注图像对应的真实值计算当前训练轮的损失函数值；
步骤S50、根据损失函数值更新待训练的旋转船只目标检测模型的模型参数，完成当前轮训练。
具体地，在进行模型训练之前，获取相关场景的船只图像数据并进行角度标注。对船只图像数据之后按照角度坐标、目标倾斜角度进行了信息标注。所使用的目标坐标及角度表达式为(x,y,w,h,θ)，其中θ采用长边定义法，即为从水平方向逆时针旋转，至第一条长边所经过的角度，取值范围为0-180°。船只图像数据使用的数据集为HRSC2016，该数据集是一个宽高比和任意方向范围较大的船舶检测数据集。该数据集包含两个场景(海上船舶和近海船舶)，具有15个对象类别。每个图像的大小范围从300×300到1500×900。该数据集有1061张图像，其中436张用于训练，181张用于验证，444张用于测试。
在步骤S20中，对每一轮训练需要采用的船只标注图像进行增强处理，增强处理的方式是将4张图片通过随机缩放、随机裁剪、随机旋转和随机排布的方式进行拼接，形成一张新的图像，即船只增强图像。其中，通过对每一种增强方式设置一个触发概率来控制增强效果。本实施例引入损失反馈机制，在进行训练的过程中，每训练一个epoch后计算一次各种损失，并评判每种损失的贡献值。如果在训练一个epoch后，角度损失贡献较少或较大(通过设置阈值进行判定)，即判断为旋转损失失衡，则在上述数据增强过程中调整旋转增强的触发概率，即将旋转触发概率设定为第一预定值，在一定程度上缓解数据不均衡问题。若判定为旋转损失不失衡，则仍按照原设定的各触发概率进行数据增强，即将旋转触发概率设定为第二预定值。示例性地，如果过低，就将旋转触发概率上调50％，过高则下调50％。这个调整比率可以实际需求设定。
如果不失衡，则当前触发概率不改变，按照当前触发概率值直接进行下一步
进一步地，待训练的旋转船只目标检测模型以YOLOv5为基础架构，在Darknet53的基础上引入CSP结构和Focus结构，形成了基础Backbone。为了更 好的提取融合特征，引入空间金字塔池化层(Spatial Pyramid Pooling layer,SPP)和FPN+PAN的结构。在SPP模块中，使用多个池化核对输入特征图进行池化操作，然后对被不同池化核处理过的特征图进行concat操作。FPN是自顶向下的结构，将高层的特征信息通过上采样的方式进行传递融合，得到用于预测的特征图。同时，一个自底向上的包含两个PAN结构的特征金字塔模块被添加到FPN后，它能够传达位置特征。FPN和包含PAN的特征金字塔结构的结合从不同的主干层对不同的检测层进行参数聚合，从而进一步提高了模型的特征提取能力。旋转船只目标检测模型还具有旋转特征对齐模块，旋转特征对齐模块通过对角度通道进行切换，采用最大池化方法评估每个特征区与目标真实区域的响应强度，再通过角度插值，得到针对角度值的特征对齐效果，能够使得后续提取到的特征更匹配旋转的船只目标。
具体来说，如图2所示，将船只增强图像输入到待训练的旋转船只目标检测模型后，进行卷积操作，构成多层卷积特征图，针对不同层级的卷积特征图都进行旋转特征对齐的操作：针对图像中的像素点，通过预设锚框匹配能够判定常规的特征提取区域。本实施例在此基础上进行改进，针对特定像素点在特定的锚框匹配状态下，切换角度通道值，针对每一个角度通道值计算对应的特征图区域，将对应的特征图区域与目标真值对应的特征图区域进行对比，得到响应大于阈值的数个方向值，并将响应最强的方向通道值作为特征对齐的对应角度。针对特征对齐后的角度信息，计算目标真值与预测结果在特征图上的偏差权重，将偏差权重融入到卷积计算中进行特征提取。特征对齐后的卷积区域的确定过程和特征提取过程参照如下公式：
Figure PCTCN2022140069-appb-000001
Y(p)＝W(r)□X(L)
其中，p为进行操作的坐标位置，w、h分别为旋转框的宽度和高度，r为卷积核半径，R(θ)为角度旋转矩阵，使用Mp操作获得与当前操作位置的真值响应最强的角度值，取最强响应角度计算得到的特征对齐后的卷积区域L。W(r)代表卷积操作，X(L)表示进行卷积操作的特征图区域，Y(p)表示提取到的特征。
进一步地，将特征对齐后的卷积区域L视为锚框，则结合真值框所覆盖区域和当前锚框所覆盖范围，可以计算两框之间的偏移量，并以它们的坐标计算得到当前预测框坐标，公式如下：
t x＝(x-x a)/w a,t y＝(y-y a)/h a
t w＝log(w/w a),t h＝log(h/h a)
t θ＝(θ-θ a)□π/180
t' x＝(x'-x a)/w a,t' y＝(y'-y a)/h a
t' w＝log(w'/w a),t' h＝log(h'/h a)
t' θ＝(θ'-θ a)□π/180
L CSL＝FL(θ,θ')
其中，x,y,w,h,θ分别表示真值框的中心坐标、宽度、高度和角度；x’,y’,w’,h’,θ’分别表示预测框的中心坐标、宽度、高度和角度；x a,y a,w a,h a,θ a分别表示锚框的中心坐标、宽度、高度和角度，符号t、t’均表示偏移量。FL表示Focal loss，L CSL表示角度分类方法本身带来的第二损失值。
进一步地，根据计算得到角度距离的权重参数调整交并比值(IOU,Intersection-over-Union)，得到第一损失值，具体的计算过程如下。
考虑到旋转目标的角度信息，如何得到精确的边框回归损失是提高检测效果的重点研究方向。目前常用的边框损失函数为IOU，GIOU，CIOU和DIOU，但它们都只适用于水平框检测，无法得到准确的旋转IOU。为了设计更好的边框回归损失，本实施例一改进了PIOU方法，以设计旋转目标检测任务的损失函数。
IOU的计算是为了评估对两个边框的重叠程度，由于旋转边界框(Oriented bounding box,OBB)和交点区域是由图像空间中的像素构成的，所以它们的面积近似为内部像素的数量。由于OBB和交集区域是由图像空间中的像素构成的，所以它们的面积是由内部像素的数量来近似的。本实施例一中将真值框和预测框分别用b、b’表示，如b即为上述公式中x,y,w,h,θ值所确定的旋转边界框，同理b’为上述公式中x’,y’,w’,h’,θ’值所确定的旋转边界框。为了判断一个点p i,j与边界框OBB之间的相对位置(内或外)，二元约束定义如下：
Figure PCTCN2022140069-appb-000002
其中，d ij为点p i,j与OBB中心坐标(x,y)之间的L 2范数距离，w,h为边界框的宽度、高度，d w和d h分别为沿水平方向和垂直方向的距离d。
设定真值框b和预测框b’的交集范围为：
Figure PCTCN2022140069-appb-000003
其并集范围为：
Figure PCTCN2022140069-appb-000004
一对边界框(b,b’)的交并比(IOU)表示为：
Figure PCTCN2022140069-appb-000005
其中PIOU大于阈值0.5时，将b,b’视为一个正样本的边界框对(这一对边界框的匹配视为正样本)。
PIOU可以很好地表示两个OBB的重叠程度，但不能计算它们的角度差。在整个旋转目标检测任务中，为了更加充分利用目标的角度信息，因此将角度信息集成到PIOU方法中可以更准确地指导回归方向。构建对角度距离敏感的权重参数，计算中引入旋转物体的角度信息，以改善PIOU损耗。该权重参数可以表示为：
Figure PCTCN2022140069-appb-000006
其中，R为真实框的长宽比，θ b和θ b’分别为真值角度以及预测角度。
故旋转感知PIoU损失，即基于角度距离的第一损失值，可以表示为：
Figure PCTCN2022140069-appb-000007
其中M表示所有为正样本的边界框对。
同时，使用特征对齐后提取的特征进行类别的训练，该步骤即通过简单的特征图匹配，使用交叉熵损失用于调整模型中对类别特征的学习，获得基于类别的第三损失值L cls(p n,t n)。
根据第一损失值、第二损失值和第三损失值构成的损失函数值如下：
Figure PCTCN2022140069-appb-000008
其中，上式中的N表示锚框的数量，obj n是一个二进制值(obj n＝1表示前景， obj n＝0表示背景，背景没有回归)。b、b’分别为真值框和预测框。θ n、θ’ n分别表示真值角度和预测角度。t n为对象的标签，p n为用Sigmoid函数计算的各类概率分布。超参数λ 1，λ 2，λ 3控制权衡，默认为{1,0.5,1}。
将获得的船只增强图像数据按照约9:1的比例分为训练集和测试集两部分，并将训练集数据以及初始参数输入到上述构建的目标检测模型中，进行前向计算。利用反向传播算法和梯度下降更新网络权重进行训练，直至网络中的损失参数收敛至获得最小误差，其中梯度优化使用Adam优化器。
进一步地，设置不同的超参数，我们将场景聚焦在无人机在近海岸的巡检作业任务上，该任务对实时性有一定的要求。本发明选取一阶段检测框架YOLOv5l作为基准，并将上述动态数据增强方法、角度分类思想以及特征对齐等方法加入到训练过程中。将本发明所涉及的模型与RoI Transformer，R2CNN等已知的先进模型进行对比，本发明设计的方法在同时考虑检测精度和检测速度的情况下，整体性能是十分突出的，其避免了角度回归的周期性问题，并且同时拥有高实时性和极高的检测精度，适合应用于类似实时性较高的巡检场景。
如图3所示，实施例二还公开了一种旋转船只目标检测模型的训练装置，该训练装置包括预处理单元100、数据增强单元200、数据输入单元300、损失计算单元400和参数更新单元500。预处理单元100用于对原始旋转船只图像进行预处理和角度标注处理，获得船只标注图像；数据增强单元200用于设定当前训练轮的旋转触发概率，按照船只标注图像进行数据增强处理，获得船只增强图像；数据输入单元300用于将船只增强图像输入到待训练的旋转船只目标检测模型中，得到预测值；损失计算单元400用于根据预测值和船只标注图像对应的真实值计算当前训练轮的损失函数值；参数更新单元500用于根据损失函数值更新待训练的旋转船只目标检测模型的模型参数，完成当前轮训练。
具体地，数据增强单元200还用于在上一轮训练完成之后，判断是否存在旋转损失失衡；若存在旋转损失失衡，则将当前训练轮的旋转触发概率设定为第一预定值；若不存在旋转损失失衡，则将当前训练轮的旋转触发概率设定为第二预定值。
进一步地，损失计算单元400还用于根据真实框所对应的中心坐标、宽度、高度、角度和预测框的中心坐标、宽度、高度、角度计算得到基于角度距离的第一损失值、基于角度分类的第二损失值和基于类别的第三损失值，第一损失值、第二损失值和第三损失值构成损失函数值。
实施例三还公开了一种计算机可读存储介质，计算机可读存储介质存储有旋转船只目标检测模型的训练程序，旋转船只目标检测模型的训练程序被处理器执行时实现上述的旋转船只目标检测模型的训练方法。
进一步地，实施例四还公开了一种计算机设备，在硬件层面，如图4所示，该计算机设备包括处理器12、内部总线13、网络接口14、计算机可读存储介质11。处理器12从计算机可读存储介质中读取对应的计算机程序然后运行，在逻辑层面上形成请求处理装置。当然，除了软件实现方式之外，本说明书一个或多个实施例并不排除其他实现方式，比如逻辑器件抑或软硬件结合的方式等等，也就是说以下处理流程的执行主体并不限定于各个逻辑单元，也可以是硬件或逻辑器件。计算机可读存储介质11上存储有旋转船只目标检测模型的训练程序，旋转船只目标检测模型的训练程序被处理器执行时实现上述的旋转船只目标检测模型的训练方法。
计算机可读存储介质包括永久性和非永久性、可移动和非可移动媒体可以由任何方法或技术来实现信息存储。信息可以是计算机可读指令、数据结构、程序的模块或其他数据。计算机可读存储介质的例子包括，但不限于相变内存(PRAM)、静态随机存取存储器(SRAM)、动态随机存取存储器(DRAM)、其他类型的随机存取存储器(RAM)、只读存储器(ROM)、电可擦除可编程只读存储器(EEPROM)、快闪记忆体或其他内存技术、只读光盘只读存储器(CD-ROM)、数字多功能光盘(DVD)或其他光学存储、磁盒式磁带、磁盘存储、量子存储器、基于石墨烯的存储介质或其他磁性存储设备或任何其他非传输介质，可用于存储可以被计算设备访问的信息。
上面对本发明的具体实施方式进行了详细描述，虽然已表示和描述了一些实施例，但本领域技术人员应该理解，在不脱离由权利要求及其等同物限定其范围的本发明的原理和精神的情况下，可以对这些实施例进行修改和完善，这些修改和完善也应在本发明的保护范围内。

Claims (10)

  1. 一种旋转船只目标检测模型的训练方法，其特征在于，所述训练方法包括：
    对原始旋转船只图像进行预处理和角度标注处理，获得船只标注图像；
    根据上一轮训练之后的旋转损失情况设定当前训练轮的旋转触发概率，按照所述船只标注图像进行数据增强处理，获得船只增强图像；
    将所述船只增强图像输入到待训练的旋转船只目标检测模型中，得到预测值；
    根据所述预测值和所述船只标注图像对应的真实值计算当前训练轮的损失函数值；
    根据所述损失函数值更新所述待训练的旋转船只目标检测模型的模型参数，完成当前轮训练。
  2. 根据权利要求1所述的旋转船只目标检测模型的训练方法，其特征在于，所述根据上一轮训练的后的旋转损失情况设定当前训练轮的旋转触发概率的方法包括：
    在上一轮训练完成之后，判断是否存在旋转损失失衡；
    若存在旋转损失失衡，则将当前训练轮的旋转触发概率设定为第一预定值；若不存在旋转损失失衡，则将当前训练轮的旋转触发概率设定为第二预定值。
  3. 根据权利要求2所述的旋转船只目标检测模型的训练方法，其特征在于，所述将所述船只增强图像输入到待训练的旋转船只目标检测模型中，得到预测值的方法包括：
    所述船只增强图像输入到待训练的旋转船只目标检测模型后，得到不同层级的卷积特征图；
    对每个层级的卷积特征图进行旋转特征对齐操作，获得特征对齐后的卷积区域；
    根据所述特征对齐后的卷积区域确定预测框的中心坐标、宽度、高度和角度，作为预测值。
  4. 根据权利要求3所述的旋转船只目标检测模型的训练方法，其特征在于，所述对每个层级的卷积特征图进行旋转特征对齐操作，获得特征对齐后的卷积 区域的方法包括：
    遍历多个角度，计算每个角度通道下，预设锚框在所述卷积特征图中确定的特征图区域与目标真值对应的特征图区域的响应值；
    根据最大的响应值所对应的角度计算得到特征对齐后的卷积区域。
  5. 根据权利要求3所述的旋转船只目标检测模型的训练方法，其特征在于，所述真实值为所述船只标注图像的真实框所对应的中心坐标、宽度、高度和角度，根据所述预测值和所述船只标注图像对应的真实值计算当前训练轮的损失函数值的方法包括：
    根据所述真实框所对应的中心坐标、宽度、高度、角度和所述预测框的中心坐标、宽度、高度、角度计算得到基于角度距离的第一损失值、基于角度分类的第二损失值和基于类别的第三损失值，所述第一损失值、所述第二损失值和所述第三损失值构成所述损失函数值。
  6. 根据权利要求5所述的旋转船只目标检测模型的训练方法，其特征在于，计算基于角度距离的第一损失值的方法包括：
    根据所述真实框所对应的中心坐标、宽度、高度、角度和所述预测框的中心坐标、宽度、高度、角度计算得到所述真实框与所述预测框的交并比值；
    根据所述真实框的角度、所述预测框的角度、所述真实框的长宽比计算得到基于角度距离的权重参数；
    根据所述基于角度距离的权重参数和所述交并比值计算得到所述第一损失值。
  7. 根据权利要求6所述的旋转船只目标检测模型的训练方法，其特征在于，所述计算得到基于角度分类的第二损失值的方法包括：
    根据所述根据所述真实框的角度、所述预测框的角度计算得到第二损失值。
  8. 一种旋转船只目标检测模型的训练装置，其特征在于，所述训练装置包括：
    预处理单元，用于对原始旋转船只图像进行预处理和角度标注处理，获得船只标注图像；
    数据增强单元，用于设定当前训练轮的旋转触发概率，按照所述船只标注 图像进行数据增强处理，获得船只增强图像；
    数据输入单元，用于将所述船只增强图像输入到待训练的旋转船只目标检测模型中，得到预测值；
    损失计算单元，用于根据所述预测值和所述船只标注图像对应的真实值计算当前训练轮的损失函数值；
    参数更新单元，用于根据所述损失函数值更新所述待训练的旋转船只目标检测模型的模型参数，完成当前轮训练。
  9. 一种计算机可读存储介质，其特征在于，所述计算机可读存储介质存储有旋转船只目标检测模型的训练程序，所述旋转船只目标检测模型的训练程序被处理器执行时实现权利要求1至7任一项所述的旋转船只目标检测模型的训练方法。
  10. 一种计算机设备，其特征在于，所述计算机设备包括计算机可读存储介质、处理器和存储在所述计算机可读存储介质中的旋转船只目标检测模型的训练程序，所述旋转船只目标检测模型的训练程序被处理器执行时实现权利要求1至7任一项所述的旋转船只目标检测模型的训练方法。
PCT/CN2022/140069 2021-12-24 2022-12-19 旋转船只目标检测模型的训练方法、训练装置和存储介质 WO2023116631A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202111598964.7 2021-12-24
CN202111598964.7A CN114419467A (zh) 2021-12-24 2021-12-24 旋转船只目标检测模型的训练方法、训练装置和存储介质

Publications (1)

Publication Number Publication Date
WO2023116631A1 true WO2023116631A1 (zh) 2023-06-29

Family

ID=81268880

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/140069 WO2023116631A1 (zh) 2021-12-24 2022-12-19 旋转船只目标检测模型的训练方法、训练装置和存储介质

Country Status (2)

Country Link
CN (1) CN114419467A (zh)
WO (1) WO2023116631A1 (zh)

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117437595A (zh) * 2023-11-27 2024-01-23 哈尔滨航天恒星数据系统科技有限公司 一种基于深度学习的渔船越界预警方法
CN117611966A (zh) * 2023-10-31 2024-02-27 仲恺农业工程学院 基于YOLOv7旋转框的香蕉识别与位姿估计方法
CN117789041A (zh) * 2024-02-28 2024-03-29 浙江华是科技股份有限公司 一种基于大气散射先验扩散模型的船舶去雾方法及系统
CN117894004A (zh) * 2023-12-27 2024-04-16 武汉科技大学 一种基于深度学习的热铸坯号识别方法及系统
CN118038452A (zh) * 2024-04-12 2024-05-14 广东海洋大学 基于弱监督的菠萝检测方法、系统及存储介质

Families Citing this family (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114419467A (zh) * 2021-12-24 2022-04-29 中国科学院深圳先进技术研究院 旋转船只目标检测模型的训练方法、训练装置和存储介质
CN114842365B (zh) * 2022-07-04 2022-11-29 中国科学院地理科学与资源研究所 一种无人机航拍目标检测识别方法和系统
CN116309696B (zh) * 2022-12-23 2023-12-01 苏州驾驶宝智能科技有限公司 一种基于改进广义交并比的多类别多目标跟踪方法及装置
CN116052094B (zh) * 2023-03-07 2023-06-09 浙江华是科技股份有限公司 一种船舶检测方法、系统及计算机存储介质

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20200143205A1 (en) * 2017-08-10 2020-05-07 Intel Corporation Convolutional neural network framework using reverse connections and objectness priors for object detection
CN113468968A (zh) * 2021-06-02 2021-10-01 中国地质大学（武汉） 一种基于无锚框的遥感图像旋转目标检测方法
CN113536936A (zh) * 2021-06-17 2021-10-22 中国人民解放军海军航空大学航空作战勤务学院 一种舰船目标检测方法及系统
CN114419467A (zh) * 2021-12-24 2022-04-29 中国科学院深圳先进技术研究院 旋转船只目标检测模型的训练方法、训练装置和存储介质

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20200143205A1 (en) * 2017-08-10 2020-05-07 Intel Corporation Convolutional neural network framework using reverse connections and objectness priors for object detection
CN113468968A (zh) * 2021-06-02 2021-10-01 中国地质大学（武汉） 一种基于无锚框的遥感图像旋转目标检测方法
CN113536936A (zh) * 2021-06-17 2021-10-22 中国人民解放军海军航空大学航空作战勤务学院 一种舰船目标检测方法及系统
CN114419467A (zh) * 2021-12-24 2022-04-29 中国科学院深圳先进技术研究院 旋转船只目标检测模型的训练方法、训练装置和存储介质

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117611966A (zh) * 2023-10-31 2024-02-27 仲恺农业工程学院 基于YOLOv7旋转框的香蕉识别与位姿估计方法
CN117437595A (zh) * 2023-11-27 2024-01-23 哈尔滨航天恒星数据系统科技有限公司 一种基于深度学习的渔船越界预警方法
CN117894004A (zh) * 2023-12-27 2024-04-16 武汉科技大学 一种基于深度学习的热铸坯号识别方法及系统
CN117789041A (zh) * 2024-02-28 2024-03-29 浙江华是科技股份有限公司 一种基于大气散射先验扩散模型的船舶去雾方法及系统
CN117789041B (zh) * 2024-02-28 2024-05-10 浙江华是科技股份有限公司 一种基于大气散射先验扩散模型的船舶去雾方法及系统
CN118038452A (zh) * 2024-04-12 2024-05-14 广东海洋大学 基于弱监督的菠萝检测方法、系统及存储介质

Also Published As

Publication number Publication date
CN114419467A (zh) 2022-04-29

Similar Documents

Publication Publication Date Title
WO2023116631A1 (zh) 旋转船只目标检测模型的训练方法、训练装置和存储介质
TWI762860B (zh) 目標檢測及目標檢測網路的訓練方法、裝置、設備及儲存媒體
WO2020108311A1 (zh) 目标对象3d检测方法、装置、介质及设备
CN113191296A (zh) 一种基于yolov5的任意朝向目标五参数检测方法
CN109145747A (zh) 一种水面全景图像语义分割方法
US20230121534A1 (en) Method and electronic device for 3d object detection using neural networks
CN113850129A (zh) 一种旋转等变的空间局部注意力遥感图像目标检测方法
CN113177503A (zh) 基于yolov5的任意朝向目标十二参数检测方法
CN115797736B (zh) 目标检测模型的训练和目标检测方法、装置、设备和介质
Zhu et al. AOPDet: Automatic organized points detector for precisely localizing objects in aerial imagery
Han et al. Research on remote sensing image target recognition based on deep convolution neural network
Song et al. Fine-grained object detection in remote sensing images via adaptive label assignment and refined-balanced feature pyramid network
US20220335572A1 (en) Semantically accurate super-resolution generative adversarial networks
Xiao et al. FDLR-Net: A feature decoupling and localization refinement network for object detection in remote sensing images
CN115100741A (zh) 一种点云行人距离风险检测方法、系统、设备和介质
CN115115601A (zh) 一种基于形变注意金字塔的遥感舰船目标检测方法
CN113128564B (zh) 一种基于深度学习的复杂背景下典型目标检测方法及系统
Qian et al. An effective pest detection method with automatic data augmentation strategy in the agricultural field
Chen et al. Orientation-aware ship detection via a rotation feature decoupling supported deep learning approach
Shi et al. RAOD: refined oriented detector with augmented feature in remote sensing images object detection
Shi et al. Obstacle type recognition in visual images via dilated convolutional neural network for unmanned surface vehicles
CN113610178A (zh) 一种基于视频监控图像的内河船舶目标检测方法和装置
Xiong et al. A unified and costless approach for improving small and long-tail object detection in aerial images of traffic scenarios
CN115035429A (zh) 一种基于复合主干网络和多预测头的航拍目标检测方法
Chi et al. Multi-target Detection for Aerial Images Based on Fully Convolutional Networks

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 22909943

Country of ref document: EP

Kind code of ref document: A1