基于CNN的机器人视觉光谱图像目标识别技术

Target recognition technology of robot vision spectral image based on CNN

传统图像目标识别技术,分配目标特征点主方向的角度较大,导致目标识别精度较差。为此,提出基于CNN的机器人视觉光谱图像目标识别技术。将光谱图像转化为灰度图,对图像进行增强处理,选取线性均值滤波,去除图像脉冲噪声和高斯噪声,将Hessian矩阵模板应用于灰度图像,比较矩形区域内每个像素,获得局部极值点,初步确定目标轮廓和边沿的空间尺度,再对目标进行一层层卷积操作,迭代训练后提取细节特征,与模板图像特征进行匹配,识别目标类别。机器人抓取带有飞机目标的视觉光谱图像,进行对比实验,结果表明,此次设计技术相比传统技术,保证了目标识别效率,在此基础上,提取特征量与模板特征量更为接近,提高了图像目标的识别精度。

In traditional image target recognition technology,the distribution angle of main direction of target feature points is large,which leads to poor target recognition accuracy.Therefore,a target recognition technology based on CNN is proposed.The spectral image is transformed into gray image,and the image is enhanced.The linear mean filter is selected to remove impulse noise and Gaussian noise.The Hessian matrix template is applied to gray image.Each pixel in rectangular region is compared to obtain local extremum.The spatial scale of target contour and edge is preliminarily determined.Then,the target is convoluted layer by layer,and the iterative training is completed to extract detailed features,and then to be matched with template image features to identify target category.The robot grabs the visual spectrum image of aircraft target,and carries out comparative experiment.The results show that,on the basis of ensuring recognition time,this design technology is closer to template feature quantity than traditional technology,and improves recognition accuracy of image target.