Abstract: An IFPN+MobileNetV2-SSD (Single Shot MultiBox Detector) model was constructed for real-time classification and detection of rubbish, in view of the problems of different sizes of detection targets and low accuracy of small target detection in rubbish classification and detection tasks. Firstly, the improved MobileNetV2 network is introduced into SSD by adding a spatial pyramid pooling module (ASPP) with hole convolution to ensure real-time network performance and accuracy while reducing the computational complexity of the network model;. Secondly, an Implicit Feature Pyramid Network (IFPN) is used to fuse SSDs from deep to shallow layers of the network step by step to detect small targets more accurately;. Finally, a focal loss function was used to adjust the weights between positive and negative samples. The experimental results show that the proposed method improves the mAP by 4.84% and reduces the detection elapsed time by 73% over the traditional SSD at a threshold value of 0.4, satisfying all the requirements of the model for edge computing devices.

Key words: Garbage Classification, Target Detection, MobileNetV2, SSD, Spatial Pyramid Pooling

摘要： 针对垃圾分类检测任务中，检测目标尺寸大小不一、小目标检测精度不高等问题，构建一种IFPN+MobileNetV2-SSD(Single Shot MultiBox Detector)模型对垃圾进行实时分类检测。首先，将改进后的MobileNetV2网络引入SSD，加入带有空洞卷积的空间金字塔池化模块(ASPP)，在减少网络模型计算复杂度的同时保证网络实时性和精确性；其次，采用隐式特征金字塔网络(IFPN)从网络的深层到浅层逐级融合SSD，更精确地检测出小目标；最后，使用focal loss函数调节正负样本之间的权重。实验结果表明，在阈值为0.4时，所提方法比传统SSD的mAP提高了4.84%，检测耗时降低了73%，满足边缘计算设备对模型的各项要求。

关键词: 垃圾分类, 目标检测, MobileNetV2, SSD, 空间金字塔池化