Study Of Detection Of Stress Cracks In Corn By Using Acoustic Analysis And Image Analysis

The stress crack percentage is an important index in evaluation of corn quality; it directly influences the evaluation of corn grade. This paper discussed the relationship between stress cracks in corn and their acoustical characteristics, analyzed the characteristics of stress cracks in corn images, and finally achieved rapid detection of stress cracks in corn. The main research contents and conclusions are as follows:1. Bringing up the experiment project. The paper summarized the present domestic and international detection methods of stress cracks in corn. According to their research results, the paper detected stress cracks in corn by using acoustic analysis and image analysis.2. Designing experiment equipment. The paper analyzed the detection principle of stress cracks in corn by using their acoustical characteristics and computer vision technique, designed impact experiment equipment and image acquisition experiment. The parameters of the impact experiment are optimized as follows: the corn impacts the under part of plate, the distance between the sound sensor and the central axis of plate is 3cm, the angle between the plate and the level is 30°, the height of the corn falls is 30cm.The optimum mode of illuminance is determined as follows: lamp-house is at the left side below the corn.3. Processing sound signal and selecting characteristic parameters. The kernel impact experiments were conducted with the glass plate, the brass plate, and the stainless steel plate. The sampling frequency of the kernel impact sound signal is 22050Hz. After filtered, endpoints detected, Fourier transformed, and power spectrum estimation, the signal was analysised in time, frequency and cepstrum domains. The paper chose ten characteristic parameters as research objects in time and frequency domains. They were signal intensity, maximum amplitude, amplitude difference, wave symmetry, decay time, signal energy, power spectrum peak value, power spectrum peak position, power spectrum x-axis centroid, and power spectrum y-axis centroid. The former five of those were time domain parameters, and the latter five of those were frequency domain parameters. After time domain analysis and power spectrum analysis, the paper came to conclusions as follows: in the experiment of big (small) kernels impacted the glass plate, signal intensity, signal energy, power spectrum peak value, and power spectrum y-axis centroid were the characteristic parameters, which could divide the intact big (small) kernels and the stress-cracked big (small) kernels; in the experiment of big kernels impacted the brass plate, signal intensity was the characteristic parameter, which could divide the intact big kernels and the stress-cracked big kernels; in the experiment of big kernels impacted the stainless steel plate, signal intensity and power spectrum x-axis centroid were the characteristic parameters, which could divide the intact big kernels and the stress-cracked big kernels. After cepstrum analysis, Mel Frequency Cepstrum Coefficient (MFCC) was picked-up to be one of the characteristic parameters, which could divide the intact kernels and the stress-cracked kernels.4. Processing corn images and extracting stress crack feature. The kernel images were acquired in image acquisition device. After image type transformation, gray-level threshold segmentation, and area revision, the kernel-area image was segmented successfully. Roberts operator, Sobel operator, Prewitt operator, LoG operator, and Canny operator were used to process the kernel images. The paper finally selected Canny operator through the comparison of edge detection results. After eliminating the two-side contour lines, the stress crack images were extracted. The number of stress cracks was counted through searching the two-dimension logic matrix.5. Recognising sound signal. Three kinds of method were used to recognize the kernel impact sound signal. (1) BP neural network was used to recognize time and frequency domain characteristic parameters. The paper came to conclusions as follows: the glass plate experiment result was best; the overall recognition rate was 71%. (2) DTW arithmetic was used to recognize cepstrum domain characteristic parameter. The paper came to conclusions as follows: the glass plate experiment result was best; the overall recognition rate was 88.50%. (3) HMM arithmetic was used to recognize cepstrum domain characteristic parameter. The paper came to conclusions as follows: this method could only divide the intact big kernels and the stress-cracked big kernels, and must use the brass plate; the overall recognition rate was 87.00%. Through the comparison of those methods, the paper came to conclusions as follows: when the experiment used the glass plate, and DTW arithmetic was used to recognize MFCC of sound signal, the division result of the intact kernels and the stress-cracked kernels was best, the overall recognition rate was 88.50%.6. Recognising corn images. The kernel images were recognized by using image analysis, and the degree of stress cracks was evaluated. The paper came to conclusions as follows: the recognition result of the multiple stress-cracked kernels was best, the recognition rate was 100%; the recognition result of the intact kernels was next best, the recognition rate was 97%; the recognition rates of single stress-cracked kernels and double stress-cracked kernels were 91% and 88%. The overall recognition rate of image analysis was 94%.7. Comparing acoustic analysis and image analysis. Detecting stress cracks in corn by using acoustic analysis and image analysis have their respective pros and cros. When the stress cracks in corn were detected by using acoustic analysis, the kernels required to divide into two kinds, big and small, each kernel impact experiment was repeated five times, the intact kernels and the double (multiple) stress-cracked kernels could be divided, the recognition rate was 88.5%. When the stress cracks in corn were detected by using image analysis, the kernels should have good transmittance of light, each kernel impact experiment was carried out one times, the intact kernels, single, double and multiple stress-cracked kernels could be divided, and the recognition rate was 94%. If both the acoustic analysis method and the image analysis method could be combined together effectively, the applied foreground of the two detection methods of stress cracks would be better.