Specific Object Detection In Satellite Remote Sensing Images Of The Yellow River Basin Based On Deep Learning

With the development of artificial intelligence technology,intelligent algorithms based on deep learning are widely used in various fields,such as object detection of satellite remote sensing images.In order to improve the clearance efficiency of "Four Disorders" problem in rivers and lakes,a real-time object detection algorithm based on deep learning can be used to improve the accuracy and speed of satellite remote sensing image recognition.YOLOv4 achieves a good balance between accuracy and speed which can meets the application’s requirements.However,this algorithm also has the problems of high computational complexity,equal integration of features from different scales and conflict between classification and location regression tasks.In this context,this paper makes some improvements on the basis of YOLOv4 and proposes an object detection algorithm of specific targets in satellite remote sensing images of the Yellow River Basin based on deep learning.The main research contents are as follows:(1)Aiming at the lack of satellite remote sensing datasets on "Four Disorders" problems in the Yellow River Basin,this paper firstly constructed the "Four Disorders" dataset,including the building class that belongs to the problem of "Unauthorized Occupation" and the greenhouses class that belongs to the problem of "Unauthorized Construction",containing 1018 images and 16012 labeled objects.Secondly,due to the relatively limited scale of "Four Disorders" dataset,the transfer learning method was used to fine-tune the model parameters by using the larger-scale DIOR satellite remote sensing dataset.(2)Aiming at the high computational complexity of YOLOv4,this paper used the spatial specificity and channel independence of Involution operator to generate the corresponding Involution kernel at each feature point and share weights in channel and proposed a feature extraction network CSPDarknet-53-Involution based on Involution,which reduced the computational cost and improved the effect of feature extraction.The experimental results on the "Four Disorders" dataset shows that,compared with the original feature extraction network,the network improves the average precision by 1.83%,reduces the number of parameters by38.96% and the computational complexity by 32.81%.(3)Aiming at the problem that the feature fusion module of YOLOv4 used equally with features of different scales which does not distinguish the fusion weights of these features,this paper proposed a feature fusion network based on weighted bi-directional feature pyramid network(Bi FPN)by bottom-up and top-down features fusion,spanning connections and weighted fusion of multi-scale features to improve the effect of feature fusion and reduce computational cost.The experimental results on the "Four Disorders" dataset shows that,compared with the original feature fusion network,the network improves the average precision by 1.42%,reduces the parameters by 33.71% and the computational complexity by 35.71%.(4))Aiming at the conflict between classification and regression tasks in the head module of YOLOv4,decoupled head was used to calculate the category,position and confidence information respectively in order to eliminate accuracy loss caused by mutual interference of the two tasks.The experimental results on the "Four Disorders" dataset shows that decoupled head improves the average precision by 0.52%.Finally,this paper conducts experiments on the method obtained by combining the above improvements and compares results with other algorithms on DIOR dataset and the "Four Disorders" dataset.The experimental results show that,compared with the original YOLOv4 algorithm,the performance of the algorithm has been improved.Among that results,the improved algorithm increases the average precision by 3.77%,reduces the parameters and computation by 17.67% and 14.79% on the constructed "Four Disorders" dataset.The improved YOLOv4 algorithm increases the average precision of the "Four Disorders" problems and reduces the computational complexity and parameters.