基于改进YOLO算法的动物源性食品检测方法

Animal Derived Food Detection Method Based on Improved YOLO Algorithm

由于动物源性食品图像的特征分布不规则,导致对其检测结果的可靠性难以得到保障,为此提出了一种基于改进YOLO算法的动物源性食品检测方法。通过YOLO V3的主干特征提取网络Darknet-53,分别对动物源性食品图像中存在的可见光和红外光进行特征提取,结合二者对应模态特征的最佳权重参数,进行特征加权融合,计算融合后特征的目标框位置损失、目标置信度损失以及类别损失,确定最终的分类。测试结果表明,设计方法对动物源性食品图像的识别结果稳定,且错误识别数量始终保持在较低水平,不受测试数据集构成的影响。

Due to the irregular feature distribution of animal derived food images,the reliability of their detection results is difficult to guarantee.Therefore,a animal derived food detection method based on improved YOLO algorithm is proposed.By using the backbone feature extraction network Darknet-53 of YOLO V3,visible and infrared light in animal source food images are extracted separately.Combining the optimal weight parameters of their corresponding modal features,feature weighted fusion is performed to calculate the target box position loss,target confidence loss,and category loss of the fused features,and determine the final classification.The test results show that the recognition results of the design method for animal derived food images are stable,and the number of incorrect recognition remains at a low level,unaffected by the composition of the test dataset.