Detection Of Stored Grain Pests Based On Image Recognition

According to China Statistical Yearbook, grain production in 2015 reached 621 million tons, achieving a continuous increase since 2003.However, in the process of grain storage, stored grain pests are becoming a great threat. In order to observe the occurrence and development of the stored grain pests in the warehouse, generally, two methods were used to estimate the density of pests, namely, sampling method and trapping method. But those two methods often involve in too much manual operation and make errors. Therefore, this paper uses image recognition method to detect the stored grain pests. The main contents are as follows:1. An image data set called LQIDSGP which contains nine kinds of stored grain pests was build. In order to achieve the real-time detection of stored grain pests, images were obtained by using the stored grain pests trapping and image acquisition system. Then a region detection method called MSERs was used to get regions. Finally, region merging was performed to get images which contain a single stored grain pest;2. The multi-object localization of stored grain pests image and image segmentation were achieved. By using prior knowledge to mark the foreground and background, the watershed algorithm was carried out to obtain images containing a single pest; on this basis, the adaptive threshold method was used to solve the problem of shadow in image segmentation;3. Recognition of stored grain pests was achieved. Standardized basic features (shape features and Hu invariant moments), color features were extracted and a recall rate of 94.7% was obtained by using SVM classifier for the nine kinds of stored grain pests;4. To verify the effectiveness of basic features, bag of features based on SURF descriptors and color features for the classification of stored grain pests, images of stored grain pests were taken by the Leica M205FA camera and an image data set (HQIDSGP) was build. Based on this, basic features, color features, bag of features were extracted. When using the bag of features alone, the recall rate reached 94.2%. When color features and basic features were used at the same time, the recall rate reached 96.7%.