一种波高信息和图像纹理特征融合的大黄鱼摄食行为量化方法

Quantitative study on the feeding behavior of Larimichthys crocea in live waters

针对实况海域大黄鱼(Larimichthys crocea)摄食行为识别受环境影响大的问题,作者以大黄鱼为对象,提出一种波高信息和图像纹理特征融合的大黄鱼摄食行为量化方法。研究基于GLCM纹理特征的提取方法,利用灰度共生矩阵计算鱼群摄食前后纹理特征的4个二次统计量(能量、相关性、对比度、逆差距)。同时考虑纹理特征受海面波高影响的问题,以小波高变化为例研究其对纹理特征的识别影响。以人工分类大黄鱼480帧图像的摄食状态为基础,判断单一纹理特征模型的识别精度,结果模型准确率达到了80%。结果表明波高在0.31 cm以上时对应的部分图像存在识别误差,误差可能性随波高增加而扩大。通过引入波高修正参数,优化单一纹理特征模型,并以另480帧大黄鱼摄食前后图像测试。海试结果说明该方法在实况海域识别准确率达到93%,能有效实现实况海域大黄鱼摄食行为量化。

This paper proposes a quantitative method on the feeding behavior of Larimichthys crocea in a live sea area to address the problem that this behavior recognition is significantly affected by the environment.This method considers Larimichthys crocea as an object and is based on the fusion of wave height information and image texture features.Research was performed on the extraction method of texture features based on GLCM using the grayscale co-occurrence matrix to calculate four quadratic statistics—energy,correlation,contrast,and inverse difference—of texture features before and after fish feeding.Texture feature methods are affected by the sea surface wave height.Therefore,the wavelet height variation example was used to study its recognition impact on texture feature methods.Artificial experience was used to determine 480 image frames,classify the feeding intensity of fish schools,and determine the recognition accuracy of a single texture feature model.The accuracy rate was 80%.The results indicate a recognition error in the corresponding part of the image when the wave height is greater than 0.31 cm.The possibility of this error increases with an increase in the wave height.The single texture feature model was optimized by introducing wave height correction parameters.Another 480 images of Larimichthys crocea were also tested before and after feeding.The sea trial results show that the accuracy of the method in the live sea area is 93%,which can effectively quantify the feeding behavior of Larimichthys crocea in the live sea area.