| China is a main agricultural country. Insect pest detection and prediction in agriculture are a main task of all levels pest control services. At the present time the prediction method of attracting pest with black light and recognizing and counting by man is generally adopted. There are some serious shortages such as bad recognition accuracy and low efficiency. It reduces seriously accuracy and timeliness of prediction and is disadvantage in guiding insect disease prevention. Therefore, this paper researched image recognition technology based on computer vision and wavelet analysis. This technology will automatically detect and predict insect pests.We designed and made the attraction device to attract agricultural pests, obtained agricultural pests' images with the color camera, processed images based on wavelet analysis. On the basis of these, we emphasized on extracting effective features, put forward recognizing pests' classes with pests' colors and texture features, and succeeded in extracting five efficient features such as color features, wave image edge moments features and so on. Then we selected features, inputted into neural network classifier, recognized pattern, presented detection results.Based on mentioned above principles, the paper finished the following contents against agricultural pests.1. To design hardware device. We attracted insect pests with common 20W black light which was used by pest control services, adjusting insect pests' posture by running water, installed color camera at proper position, and designed illuminating system so that we could obtain distinct insect pest image. Finally, insect pest image was inputted into the computer by color image acquisition card.2. Wavelet transform. We researched wavelet transform in detail, programmed software with Visual C++6.0, and decomposed and composed the image under three scales by wavelet transform.3. To preprocess the image. We fulfilled image enhancement with wavelet coefficient to denoise after we decomposed the image by wavelet.4. Feature extraction. This step was a key procedure in recognizing the insect pest image. We extracted the following features: (1) Morphologic features. Area and perimeter were extracted after the image was changed the gray image and the two-value image. (2)Color features. Color model space was changed from RGB(red, green, blue) to HSI (hue, saturation, intensity) space. According to H value, red, yellow, cyan, blue,purple, and black, up to seven colors, were determined. Area, perimeter, width, height, roundness, up to five morphologic features, were extracted to each color. (3)Wavelet scale invariant moments features. Seven invariant moments were calculated with wavelet coefficient under various scales. (4) Wavelet image edge moments. Multiple-scale edge detection was completed and image edge moments were calculated by utilizing wavelet multi-resolution character. (5) Fractal dimension features. They were calculated by energy ratio under various scales.5. Feature selection. To all features extracted, we selected 12 features, for example, area, perimeter, six color features, two image edge moment features, and two fractal dimension features. Considering practical situation, we contracted six color features to one dimension. Therefore, we obtained seven efficient features in all.6. To recognize and classify agricultural pest. We designed BP neural network and trained it by seven efficient features extracted associated with object of study. We analyzed various improved methods because of the problem that conventional BP algorithm was liable to trap in local minimum value and was slow rate of convergence.We implemented mentioned above functions with Visual C++6.0 language, developed software package, associated with designed hardware system, tested 8 kinds of common agricultural pest, for example, Eterusia aedea linneus, Parasa consocia, Marumba aperchius, asparagus caterpillar, Maize borer, cotton bollworm, army worm, and so on. Rate of recognition is up to 85.7%.When it c... |
PDF Full Text Request