Research On Text Detection Algorithm Of Agricultural Materials Image Based On Segmentation

The text information in agricultural materials image provides an important basis for consumers to purchase agricultural material,and also helps agricultural materials safety supervision and enforcement departments to detect and analyze potential agricultural materials safety problems.The detection of text in agricultural materials image can help agricultural safety supervision and enforcement departments to improve the manual paper record supervision method and enhance the efficiency of law enforcement,therefore,agricultural materials image text detection has greater research significance for agricultural safety supervision and identification of agricultural content.However,there is no publicly available dataset of agricultural materials image,and the detection of text in agricultural materials image faces many challenges,such as complex image backgrounds,image distortion,text shapes and sizes,and the existing algorithms have yet to improve the effectiveness of text detection.To address the above problems,the main research of the thesis is as follows:(1)To address the lack of agricultural materials image text detection dataset,the thesis constructs a dataset containing 708 images and 11322 text boxes,labeled as a polygon formed by a number of coordinate points connected clockwise,and divides the dataset with the self-service sampling method.(2)To address the accuracy problem of agricultural materials image text detection,the thesis proposes an attention mechanism-based text detection algorithm for agricultural materials image.The thesis selects a backbone network with stronger feature extraction ability,and designs a dual feature fusion module to integrate local and global contextual feature representations to enhance the extraction of text features at different levels,and finally generates predicted text boxes by post-processing with a scaling extension algorithm.The final detection result is obtained.The algorithm can accurately frame the agricultural materials image text regions,and the accuracy,recall and F-score on the agricultural materials image dataset are 91.4%,87.3% and 89.3% respectively,which are higher than the comparison algorithms.(3)To address the problem of fast agricultural materials image text detection,the paper proposes a Ghost module-based algorithm for agricultural materials image text detection.The paper uses a lightweight network with lower parameters to extract the underlying features,introduces a multi-scale feature fusion module to obtain feature fusion between multiple layers,and uses a differentiable binary post-processing algorithm to predict the text,enabling it to detect text in agricultural materials image quickly.Experimental results show that the accuracy of the algorithm on the agricultural materials image dataset is basically at the level of mainstream algorithms,with a detection speed of18.6 FPS and a covariance of 2.99 M.The algorithm can be ported to mobile and detected quickly,and the paper finally deploys the algorithm to mobile devices and runs successfully.