SOLO-DAFF:一种面向肉牛体尺测量的图像实例分割算法

SOLO-DAFF:an image instance segmentation algorithm for beef cattle body size measurement

为了解决复杂背景下因颜色区分度小造成肉牛图像分割精度较低问题,试验采用SOLO-DAFF模型对肉牛图像进行实例分割,即在RGB彩色图像中引入深度信息,提升信息通量,增大目标与背景间的特征差异;通过在残差网络(residual network, ResNet)中增加基于卷积块的注意机制(convolutional block attention module, CBAM)提升目标特征提取精度;在特征金字塔网络(feature pyramid networks, FPN)中结合基于压缩和激发(squeeze-and-excitation networks, SE)的注意力机制进行特征图融合,得到包含准确空间位置信息的特征图;通过在分割网络头部的掩模损失函数中融入平滑系数(α),实现对肉牛图像的像素级实时分割。使用ZED双目相机采集110头肉牛的图像(共计1 500张肉牛的RGB彩色图像及相对应的深度图像),在此肉牛图像数据集上利用SOLO-DAFF模型、SOLO v2模型等进行肉牛图像实例分割验证试验。结果表明:SOLO-DAFF模型对肉牛图像的分割精度相较于SOLO v2模型提高了5.56百分点,证明了SOLO-DAFF模型有效,且SOLO-DAFF模型改善了肉牛细节部位的分割效果。说明SOLO-DAFF模型可实现从肉牛图像中获取精确的肉牛区域,提高了肉牛图像实例分割精度。

In order to solve the problem of low segmentation accuracy of beef cattle image due to small color differentiation in complex background,SOLO-DAFF method was used to segment instances of beef cattle image,that is,depth information was introduced into RGB color image to improve information throughput and increase the feature difference between target and background.By adding the convolutional block attention module(CBAM)to the residual network(ResNet),the accuracy of object feature extraction is further improved.In the feature pyramid network(FPN),the attention mechanism based on the squeeze-and-excitation networks(SE)was combined to fuse the feature map,so as to obtain the feature map containing accurate spatial location information.By incorporating the smoothing coefficient(α)into the mask loss function of the segmentation network head,the real-time segmentation of beef cattle image at pixel level was realized.The images from 110 beef catle were collected by a ZED binocular camera(a total of1500 RCB color images and corresponding depth images of beef cattle),and the beef cattle image instance segmentation experiment was carried out on the self-made beef cattle image data set by using SOLO-DAFF model and SOLO v2 model.The results showed that the segmentation accuracy of the SOLO-DAFF model for beef cattle images was improved 5.56 percentages compared with the SOLO v2 model,which proved the effectiveness of the SOLO-DAFF model.SOLO-DAFF model improved the segmentation effect of beef cattle details.The results indicated that the SOLO-DAFF model could achieve the collection of accurate beef cattle region from the beef images,and improve the segmentation accuracy of beef cattle image instance.