Age reading of sardine scale using neural network

1.利用人工神经网络建立了一个日本沙丁鱼鳞片年循环计数系统。利用图像处理技术对一个尺度的视频图像进行处理，并存储五个径向断面的光强。确定强度的峰值和位置沿着横断面，并将其归一化为输入到神经网络的5^<**>40目二维数据.年龄测定的系统的性能进行了检查，使用的数据集从20个样本的沙丁鱼鳞片。在学习过程中，神经网络表现出了从训练样本中确定年龄的学习能力，没有错误。然而，训练后的神经网络对测试样本的正确识别率不超过50%，错误率较低.以下方面有待审查：（1）由于学习数据集的数量太少，需要更多的学习，（2）照明还有改进的空间，以获得具有清晰的年循环的更好图像，（3）由于径向凹槽的对比度比年循环更高，在刻度图像中阻碍了对年循环的正确计数，需要对图像进行一些预处理过程以消除噪声凹槽，以及（4）对于大于5 × 40的测量断面数量和/或大于40的网格尺寸，应当进一步检查神经网络的输入数据，这一方面增加了分辨率，另一方面增加了神经网络中的单元数量，即更多的计算时间。

1. A personal-computer-based system was developed for counting annual circuli in scales of Japanese sardine using artificial neural network. The image processing was used to digitize the video image of a scale and store the light intensity from five radial transects. The values and positions of the peaks of the intensity along the transect were determined and normalized to input to the neural network as 5^<**>40 mesh two-dimensional data.2. Performance of age determination of the system was examined using a data set from 20 samples of sardine scales. In the learning, the neural network showed learning ability to determine age from training samples without error. However, the performance of the trained neural network for test samples got no further than 50% of correct recognition rate, while error rate was low.3. The followings remain to be examined : (1) much more learning is needed as the number of learning data set were too small, (2) there is room for improvement of illumination to get better image with clear annual circuli, (3) since radial grooves, which have higher contrast than annual circuli in the scale image prevent counting annual circuli correctly, some pre-processing procedure of the image is required to eliminate the noisy grooves, and (4) input data to the neural network should be examined further for larger number of measureing transect and/or larger mesh size than 5^<**>40, which increase resolution power on the one side and the number of units in the neural network on the other hand, i.e. much more calculating time.