基于背景饱和度压缩与差异阈值分割融合的香蕉串识别

Bunch of bananas recognition based on background saturation compression and segmentation&fusion with different threshold range

针对卷积神经网络识别香蕉串样本需求量大、训练时间长、硬件配置高,而传统图像处理方法易受光照、背景颜色影响导致识别的香蕉串不完整且噪声点多等问题,提出一种基于图像背景饱和度压缩与差异阈值分割融合的香蕉串识别方法。首先根据饱和度图像中灰度等级的像素比例自适应确定香蕉串区域的高、低饱和度阈值,然后对小于低饱和度阈值的图像背景做伽马变换;对大于高饱和度阈值的图像背景做半值压缩,进而增强香蕉串区域与环境背景的对比度。接着,采用大、小阈值范围分别对饱和度分量与色调分量的差值图像进行分割,并对分割结果进行孔洞填充和连通域提取,将获取的背景噪声与大、小阈值范围的分割结果做差值融合进一步去噪,从而得到噪声点少、准确度较高的香蕉串。试验结果表明,对自然香蕉园环境下采集的图像样本香蕉串识别的准确度高于0.85的占比39.29%,介于0.80~0.85的占比46.43%,低于0.80的占比14.28%。本文方法能较好地适应不同光照和环境颜色下香蕉串的识别。

To solve the problems that convolution neural network has some disadvantages in bunch of bananas recognition,including large scale training samples,long training time,high hardware requirements,and that traditional image processing methods are easily affected by illumination and background color,which results in recognizing an incomplete bunch of bananas with many noise points,a bunch of bananas recognition method based on image background saturation compression,different threshold segmentation,and fusion is proposed.Firstly,the high and low saturation thresholds of bunch of bananas potential regions are adaptively determined according to the pixel proportion of the grayscale in the saturation image.Then,gamma transform is performed on the image background saturation,which is less than the low saturation threshold,and half value compression is performed on the image background,which is greater than the high saturation threshold,to enhance the saturation contrast between bunch of bananas potential region and environment background.Then,the difference image of saturation component and hue component is segmented by using large and small threshold range,respectively.Environmental background noise is extracted from the threshold segmentation result by hole filling and connected domain extracting.The segmentation result of large and small threshold range is the difference fused for reducing background noise points to get the bunch of bananas with higher accuracy and fewer noise points.The experimental results show that the accuracy of bunch of bananas recognition higher than 0.85 is accounted for 39.29%,between 0.80 and 0.85 for 46.43%,and less than 0.80 for 14.28%with the images sampled in a natural banana plantation environment.This method can adapt to the bunch of bananas recognition under different light and environmental colors.