Deep Learning Based Recognition And Detection Method Of Watermelon Leaf Diseases

Diseases are one of the main problems affecting the yield of watermelon.The occurrence of watermelon during the whole growth period will seriously affect its quality and yield.Therefore,how to quickly and accurately identify and detect watermelon diseases and prevent them effectively is particularly important for watermelon crop production.At present,the identification of watermelon diseases generally requires experts to conduct on-site observation or take samples back to the laboratory for analysis.This method is not only time-consuming and labor-intensive,but also not conducive to promotion.Due to the problems that watermelon diseased leaves are not easy to flatten,easily cover each other,and the area of diseased spots is too large or too small,traditional watermelon disease identification and detection methods are not only low in accuracy but also weak in robustness.In recent years,deep learning algorithms have been widely used in the field of computer vision,and have achieved good results in many large-scale recognition and detection tasks,and have gradually become the preferred solution in the field of agricultural disease recognition and detection.In this paper,based on the deep learning algorithm,the target recognition and detection methods and theories of Watermelon Leaf Disease image are systematically studied.Combined with the three-dimensional modeling,generation of confrontation network and other algorithms,the optimization methods of self detection network are improved,and the feasible new method of Watermelon Leaf Disease Recognition and detection task is explored.The main work of this paper is divided into the following aspects:(1)Construct a watermelon leaf disease data set.After several years of field investigations,three different data collection methods are adopted,including single leaf shooting on the spot,leaf picking and shooting in the field,and sampling back to the laboratory for shooting.After data cleaning,experts in pests and diseases After identification and data labeling,the watermelon leaf disease and insect pest data set WDPD2020 and Hunan common watermelon disease data set HWCDD2020 were constructed.There are 4032 photos of watermelon leaves,including 9 kinds of diseases and 5 kinds of insect pests,which are divided into 63 categories according to the actual situation of each leaf.(2)A method based on three-dimensional modeling for disease recognition of non occluded watermelon leaves was proposed.The three-dimensional information of the watermelon leaves is obtained by continuously shooting 15 times of the watermelon leaves in a certain sequence,and the three-dimensional model of the watermelon leaves is established based on this,and then it is rotated in 4 ways in turn,and for each watermelon leaf threedimensional The model intercepts 96 pictures under different rotation angles,and adds them to the HWCDD2020 data set for training and testing.Compared with the method without 3D modeling,under different training modes,the overall accuracy is improved by 1.8%-5.4%,and the kappa coefficient is improved by 0.02-0.08.The results show that this method can effectively improve the recognition effect of unshielded watermelon leaf diseases.(3)A method of Watermelon Leaf Disease Recognition Based on generation antagonism network was proposed.By improving the Edge-connect network,constructing the occluded watermelon leaf disease image prediction model,and repairing the occlusion randomly generated in the test set of HWCDD2020,thereby improving the recognition rate of occluded watermelon leaf diseases.Compared with the occluded watermelon diseased leaves,the overall accuracy of the repaired watermelon diseased leaf images in different models has been improved by 0.6%-28.5%,and the kappa coefficient has been improved by 0-0.35.The results show that this method can effectively improve the recognition effect of shielding watermelon leaf diseases.(4)A method of Watermelon Leaf Disease Detection Based on improved SSD network was proposed.According to the actual characteristics of watermelon diseased leaves,this method proposes a new type of pre-selection box equation.At the same time,the relevant formula of the pre-selection box setting is optimized to adapt to different sizes of SSD networks.The method proposed in this article is used in SSD300 networks and multiple SSDs.Experiments in derivative networks show that in SSD networks of different scales,compared with the default values of SSD networks,the average accuracy of the method proposed in this paper is improved by 1.1%-3.8%,and the average cross-to-parallel ratio is improved by 0.7%-1.2%.Compared with the empirical values on the Internet,the average accuracy is improved by 0.7%-3%,and the average cross-to-bin ratio is improved by 0.8%-1.6%.The results show that this method effectively improves the overall effect of watermelon leaf disease detection.