Study On Key Frame Extraction Method For Identification Of Greenhouse Vegetable Foliage Diseases

When diagnosing the greenhouse vegetable diseases, there still exist problems that the applicability of advanced technology is relatively low and farmers lack accurate diagnosing tool, who have to rely on there own experience and guidance of agriculture expert. Therefore, it is of great significance to improve the diagnosing accuracy and offer an easy to use, realtime and reliable diagnosing tool to farmers.This paper takes the monitoring videos as the data source and put forward a real-time diagnosing method for greenhouse vegetables diseases based on monitoring video features. This paper establishes the video capture model and key frame extraction model by research on computer vision and video processing, which avhieves the key frames containing disease information from video sequence. This paper establishes the disease spots segmentation model and disease diagnosing model by research on image process and machine learning to avhieve the realtime diagnosis of greeshouse vegetable diseases. The main achievements of this study are as follows:(1) A monitoring video capture method orienting the identification of greenhouse vegetable diseases was presented in this paper according to the characteristics of the diseases of the greenhouse leafy vegetables. The monitoring video capture method utilized the technology of IOT and the information acquired by sensors and monitoring cameras in the greenhouse. The greenhouse video acquiring method fused the case based reasoning and fuzzy reasoning, more specially; monitoring videos can be acquired through the greenhouse video acquiring method by matching the data collected by sensors in real time with the combination of environmental conditions in the knowledge base. The fusion of case based reasoning and fuzzy reasoning can not only make up the incompleteness of the use with single method of case retrieval, but also ensure the accuracy of the data. The results showed that the recall of the monitoring video capture method was 95.4%, which indicated that the monitoring video capture method can meet the video data requirement of identification of greenhouse vegetable diseases.(2) A method that combines the visual saliency and online clustering to extract the key frame from greenhouse vegetables monitoring video was presented in this paper. Firstly X2 histograms are used to measure the similarity of each frame to the first frame, which eliminates the meaningless frames and improve data processing efficiency and costs. Then, all frames will be converted to HSV color space and a saliency map of each frame is generated based on H component value and S component value. According to the saliency map, the salient region can be obtained. During the process of extracting the salient region, there is a possibility that the information of disease spots is lost. Therefore, morphological method would be utilized to restore the lost information. Finally, online clustering is performed to classify the salient regions into different clusters, and mean pixels value is used to select the key frames. The results indicate that this method can obtain information of entire leaf area of vegetables and extract the key frame effectively.(3) An Image Segmentation Method for Processing Greenhouse Vegetable Disease Spots was presented in this paper to reduce the impact of strong illumination and complicated background on the segmentation of greenhouse vegetable disease spots. Firstly, Wapper and filter method was adopted to select the features. Then, a decision tree model was built by CART method with the selected features, which would be used to segment greenhouse vegetable disease spots. Taking K-means method and OTSU method as comparisons, this research conducted the evaluation of the proposed method. The results reach an accuracy of 94.56%, which indicate that the proposed method could achieve the disease spots effectively regardless of the strong illumination and complicated background.(4) A solution of extracting and selecting features was presented in this paper.25 features were extracted after acquiring the disease spots image. This paper reduce the number of features by attribute reduction bases on genetic algorithm, which eliminates the meaningless features and avhieve the representative feature sutset. The solution lows the complexity of the classifer and increase its efficiency. The results show that the method manages to reduce the 25 features to 12 features.(5) A classifer selection experiment was presented in this paper. BP neural network based classifer, SVM of different kernel functions based classifer and decision tree based classifer were established in this paper. Then the classifers were trained and tested by the data. The results showed that the radial and function based SVM classifer with optimized parameters achieved the best accuracy, which was 90%.(6) Designed and developed real-time diagnosing system for greenhouse house vegetable diseases. This paper designed and developed the whole system, based on the study on theory and methods, to offer an easy to use, realtime and reliable diagnosing tool to famers.