Abstract: In mining operations, metal objects such as anchor bolts, anchor cables, excavator teeth, and pallets often get mixed into the raw coal conveyor belts. These foreign objects must be removed by iron separators to prevent collisions or punctures that could disrupt the normal operation of the conveyor belts. A foreign object recognition method suitable for low illumination and dust-fog environments is proposed for iron separators , which often encountered in mining belt conveyors. First, abnormal and normal images from iron separators on open-pit coal mine conveyor belts were collected. Contrast Limited Adaptive Histogram Equalization was applied to pre-process low-illumination images, enhancing image contrast and clarity. To simulate real dust-fog conditions, a random fogging method was employed, improving the model's generalization ability. Then, using the transfer learning based EfficientNet-B2 network, incorporating multiple mobile inverted bottleneck convolutional modules into the network architecture. This allowed for the stacking and analysis of feature maps at different levels, extracting deep feature signals from the images. The high-dimensional feature maps were reduced to low-dimensional vectors through global average pooling, and the final image classifications—qualified or abnormal—were output through a fully connected layer. The experimental dataset comprises 3 000 images of iron separators collected from an open-pit coal mine and 600 fogged images. The proposed foreign object monitoring algorithm was applied to monitor the iron separators on the conveyor belts of an open-pit coal mine. Comparative experiments were conducted, and results show that the proposed model achieves faster stable iterations and lower loss values, outperforming other existing convolutional neural network models across various performance metrics. Specifically, it achieved an accuracy of 99.79%, precision of 99.07%, recall of 99.01%, and F1-Score of 0.990 4. These results indicate that the model can accurately and effectively classify the adsorption states of iron separators.