FMNet: Semantic Segmentation of Small Targets in Food Factories by Fusing Multi-Scale Features

How to better integrate features of different scales is an important research direction in the field of semantic segmentation. In the task of semantic segmentation of small objects, there is often a problem of incorrect object boundary segmentation, which is caused by the model's inability to effectively integrate multi-scale information and obtain sufficient features. To solve this problem, we designed a fusing multi-scale features model.It is included by multi scale pyramid pooling(MSPP) module and atrous convolution fusion(ACF) module.MSPP module is designed to extract multi-scale features based on spatial pyramid pooling with different rates.ACF module has different rates of dilated convolution. These two modules together form an overall fusing multi-scale module. We created a dataset for online detection in a food workshop and our model outperforms other semantic segmentation models on this dataset, with an accuracy of up to 97%. In addition, the computational and parameter costs of our model only slightly increase.