The wear state of abrasive belt is one of the important factors affecting the grinding precision of belt grinding processes. Accurate monitoring of abrasive belt wear can not only provide the basis for accurate material removal model to improve grinding accuracy, but also can replace the belt to avoid surface burn in time. However, most of the existing abrasive belt wear monitoring methods are only suitable for monitoring the belt wear state under specific grinding parameters, are not universal. This paper introduces a method of belt wear state monitoring based on machine vision and image-processing. All the surface images of the belt were obtained from the new belt to the worn-out of the belt by a non-contact electron microscope. The features of abrasive belt surface images are extracted from RGB color space and wavelet texture. By analyzing the trendency of the extracted features in the whole grinding process, the wear state is divided into three categories. Three image features related to the wear state are selected: the first order distance of color component R, the entropy of horizontal subgraph, and vertical subgraph of texture feature. Based on the selected features and the random forest classification algorithm, the wear state classifier of abrasive belt is established. The performance of the classifier is verified and evaluated by using the data subset of different images. The results show that the proposed method has high recognition accuracy for the belt wear state, and the accuracy can reach 99% in the accelerated wear stage. The proposed method is suitable for the monitoring of the belt wear state by the surface images of the abrasive belt measured under different grinding parameters and different measurement parameters.