A robust detection method for illegal occupation of farmland in national territory based on improved SSD
The third National Land Survey indicates the actual farmland area in China is 1.918 billion acres,113 million acres less compared with the previous survey,moving closer to the alert level of 1.8 billion acres of farmland.Illegal occupation of farmland still exists in China,leading to the reduction of farmland area.Against such a backdrop,a rapid and accurate detection of illegal occupation is the key to protecting China's farmlands and safeguarding the alert level of farmland.The current detection methods for illegal occupation usually have difficulty in balancing accuracy and real-time stability.To address this,we combine high-definition videos with deep learning,relying on real-time video data collected by high-definition cameras mounted on communication towers of China Tower Corporation,and propose an improved SSD(single shot multibox detector,SSD)algorithm for real-time intelligent detection of illegal occupation of farmland.Our algorithm achieves a higher accuracy in real-time performance and also addresses other problems of SSD algorithm in image processing.First,we build a dataset of illegal occupation behaviors of farmland.Cameras mounted on iron towers across Chongqing are employed to obtain real-time video data,which are converted into image data through frame extraction.Samples containing illegally occupied objects are selected.Retinex image dehazing algorithm is used to dehaze images.Segmentation and cropping methods are introduced to generate the same-sized images.Then,data augmentation is performed on the images to expand the dataset and they are labled to obtain the final dataset of illegal occupation behaviors for deep learning model training,consisting of 5 574 images in 15 categories,laying a foundation for subsequent algorithm applications.Next,we conduct deep learning for detecting illegal occupation of farmland.The SSD object detection algorithm is employed and the issues including high proportions of small objects,low detection accuracy,and poor real-time stability are targeted in the detection process of illegal occupation behaviors of farmland.The following improvements are made on the SSD algorithm:ResNet-50 is employed to replace the backbone network to enhance the feature extraction capability of algorithms for illegally occupied objects and effectively address gradient vanishing and exploding that occur during model training;the feature fusion module(R50-FFM)is built to fuse high and low layers to enhance the detailed features of small objects and improve the detection of small objects;the short-cut module is built to generate the prediction feature pyramid of SSD;the K-means clustering algorithm is used to optimize the aspect ratio of the candidate bounding box and make it more suitable for the self-made dataset;the matching probability of the objects and the candidate bounding box is improved and the detection accuracy is further enhanced;the Joint Weight Adjustment Module(JWAM)is created to enhance useful object information,suppressing irrelevant information and focusing more on the useful information of small objects in illegal occupation behaviors.Finally,8 580 samples of 15 object categories of illegal occupation of national farmland are taken as experimental objects.Our experimental results show the improved SSD object detection algorithm has a mean average precision(mAP)of 91.24％in detecting illegal occupation objects,which is 5.74％higher than that of the SSD algorithm.It performs better than many other object detection algorithms and effectively enhances the detection efficiency of small objects in illegal occupation behaviors.It also meets the pressing needs of real-time detection of illegal occupation of farmland and thus serves as a reference for monitoring lands and resources.
国家科技期刊平台
NSTL
万方数据
