运输皮带在煤矿运输过程中起着重要的作用，一旦发生破损，不但会降低生产效率，还会带来经济损失，严重时会威胁操作人员的安全。所以，运输皮带损伤检测对煤矿企业的生产安全起着至关重要的作用。语义分割在运输皮带损伤检测领域颇具优势却少有研究，因此本课题具有重要研究意义和实用价值。本文的主要工作如下：

   （1）针对卷积神经网络（Convolutional Neural Network，CNN）对细微损伤区域无法有效分割的问题，设计了一种高精度的语义分割模型，能将CNN与Transformer模型特征进行融合。该模型的主干网络用作局部特征和全局特征的提取、经过特征聚合、增强特征、统一尺寸；最后融合所有特征并经上采样，完成损伤分割。另外还设计了一种联合损失函数，由交叉信息熵和Dice损失函数构成。用该损失函数设计的模型与经典模型相比，平均交并比和平均准确率分别至少提升2.84%和2.42%；与最新模型相比，平均交并比和平均准确率分别提升了2.16%和1.87%。利用设计的模型与损失函数，可以在视觉上明显辨别出细微损伤，这是其他方法不能做到的。

   （2）针对语义分割模型无法直接部署边缘计算设备的问题，设计了一种轻量级语义分割模型。该模型利用MobileFormer标准块构建了2种不同参数量的轻量级主干网络，并以其作为编码器，以轻量级空洞池化金字塔模块为解码器。在7种模型的对比实验中，当参数量较少和较多时模型的计算复杂度均为最小，分别为0.517G、2.872G FLOPs。在参数量较少的模型中，损伤分割准确率和交并比分别提升了3.46%和8.64%；在参数量较多的模型中，损伤分割准确率和交并比都至少提升了1.64%和11.17%。另外，在视觉上也能最容易辨别出皮带上细微的损伤。可见，所设计的轻量级模型推理速度最快，准确分割损伤的能力最强。

   Conveyor belts play a crucial role in the coal mine transportation process. If they are damaged, it not only reduces production efficiency but also brings economic losses, and in severe cases, it may even threaten the safety of the operators. Therefore, the detection of conveyor belt damage is of vital importance to the production safety of coal mining enterprises. Semantic segmentation has significant advantages in the field of conveyor belt damage detection but has been little studied, thus this research topic holds important significance and practical value. The main work of this paper is as follows:

    (1) In response to the problem that the convolutional neural network (CNN) cannot effectively segment subtle damage areas, a high-precision semantic segmentation model has been designed that fuses features from both CNN and Transformer models. The backbone of this model is used for the extraction of local and global features, followed by feature aggregation to enhance and unify the feature sizes. Finally, all features are fused and upscaled to complete the segmentation of damage. Additionally, a combined loss function has been designed, composed of cross-entropy and Dice loss functions. The model designed with this loss function, compared to classical models, has at least a 2.84% improvement in mean intersection-over-union (IoU) and a 2.42% improvement in mean accuracy; when compared to the latest models, it has a 2.16% improvement in mean IoU and a 1.87% improvement in mean accuracy. Utilizing the designed model and loss function, subtle damages can be visually distinguished, which is something other methods cannot achieve.

    (2) To address the issue that semantic segmentation models cannot be directly deployed on edge computing devices, a lightweight semantic segmentation model has been designed. This model utilizes MobileFormer standard blocks to construct two lightweight backbone networks with different parameter sizes, serving as the encoder, and employs a lightweight atrous spatial pyramid pooling module as the decoder. In comparative experiments among seven models, the computational complexity of the model is the lowest when the parameter size is both small and large, with 0.517G and 2.872G FLOPs, respectively. In the model with fewer parameters, the accuracy and IoU for damage segmentation have been improved by 3.46% and 8.64%, respectively; in the model with more parameters, the accuracy and IoU for damage segmentation have been improved by at least 1.64% and 11.17%. Additionally, the model visually distinguishes subtle damage on the conveyor belt with the greatest ease. It is evident that the designed lightweight model has the fastest inference speed and the strongest capability for accurately segmenting damage.