基于改进的更快的卷积神经网络特征区域的淡水鱼鱼鳃切口点定位

为了提高鱼产品加工过程中鱼鳃切口点定位的准确度,采用改进的更快的卷积神经网络特征区域(faster convolution-al neural network feature region,Faster RCNN)对淡水鱼的鱼鳃部位进行检测和定位.首先,为了增强主干网络VGG16的特征提取能力,加入批归一化(batch normalization,BN)层对其进行结构优化,提高了网络识别的准确率.其次,当物体处于预设的交叉阈值范围时,非最大值抑制(non-maximum suppression,NMS)算法存在目标漏检的问题.采用Soft-NMS算法替代NMS算法,增强了目标检测的性能.通过在淡水鱼数据集进行的实验结果表明,改进的Faster RCNN网络对鱼鳃切口定位准确率达到了96％,较未改进网络提高了6％,为后续生产线中鱼鳃的精准切割奠定了基础.

RESEARCH ON LOCATION OF FRESHWATER FISH GILL CUT POINTS BASED ON IMPROVED FASTER RCNN

In order to improve the accuracy of the positioning of fish gill cut points during the processing of fish products, this paper uses an improved Faster RCNN (faster convolutional neural network feature region) to detect and locate the gills of freshwater fish. First, in order to enhance the feature extraction capability of the backbone network VGG16, the BN layer is added to optimize its structure, which improves the accuracy of network recognition. Secondly, when the object is in the preset crossing threshold range, the NMS (non-maximum suppression) algorithm has the problem of missed target detection. This paper uses Soft-NMS algorithm to replace NMS algorithm, which enhances the performance of target detection. The experimental results conducted on the freshwater fish dataset show that the improved Faster RCNN network has achieved 96% accuracy in positioning gill incision, which is 6% higher than the unimproved network, laying the foundation for precise cutting of fish gills in subsequent production lines.