基于SSD算法的自然条件下青苹果识别

Recognition of green apple in natural scenes based on SSD algorithm

针对自然条件下青苹果与背景颜色相似而识别困难的问题,提出了一种SSD深度学习算法与相关图像处理算法相结合的方法。首先采用SSD算法检测图像中的青苹果,将苹果区域用边界框标记并得到其位置坐标;然后根据该坐标将非苹果区域像素用黑色填充,仅保留边界框内的像素;最后运用k-means算法、RGB空间中的R-B色差法及YUV颜色空间中的U分量阈值分割法识别边界框内的苹果轮廓,通过对比3种方法的识别效果得到最优的苹果识别方法。实验表明,SSD算法与U分量阈值分割法相结合的方法识别精度最高,平均分割误差为8.04%,假阳性率为0.55%,假阴性率为9.05%,叠加系数为88.60%。

To solve the problem that colors of green apple and background are too similar in natural scenes to identify,a method combining SSD deep learning algorithm and related image processing algorithm is proposed.Firstly,the green apple in the image was detected by SSD algorithm,and the apple area was marked by a rectangular box with its regression position coordinates.Secondly,according to the coordinates of this position,only the pixels in the rectangular box were preserved,and all other non-apple areas were removed.Finally,k-means algorithm,R-B chromatic aberration method in RGB space and U component threshold value in YUV color space were used to identify the fruit area in the rectangular box respectively.And the optimal fruit recognition method was obtained by comparing the recognition effects of the three methods.Experiment results show that SSD algorithm combined with U component threshold segmentation has the highest recognition accuracy,with an average segmentation error of 8.04%,a false positive rate of 0.55%,a false negative rate of 9.05%and a stacking coefficient of 88.60%.