Abstract: The working environment of coal mine belt conveyors is complex, and it is difficult to effectively distinguish foreign objects from background noise in environmental images. The extraction rules that rely on fixed features are not suitable for diverse forms of foreign objects, further increasing the difficulty of foreign object recognition. Therefore, with the aim of improving the efficiency and stability of coal mine belt conveyors, this paper proposes a conveyor foreign object recognition method based on RetinaNet deep learning. Firstly, the structure of the RetinaNet deep learning model is analyzed, and a transport plane sample candidate region is established by combining the cross entropy loss function. The RetinaNet deep learning algorithm is used to classify the samples. Through a multi-level convolutional structure, RetinaNet can capture the detailed features of foreign objects and automatically extract multi-level features of foreign objects from complex backgrounds. Based on this, by introducing weight coefficients, difficult to distinguish samples and easy to distinguish samples can be distinguished. Then, high-frequency and low-frequency features of the sample are output through convolution and average pooling operations. Afterwards, establish a transport aircraft foreign object recognition box, input sample features into it, calculate the confidence of the recognition target, derive the deviation function, and provide the loss function of the height, width, and volume features of the foreign object target. Finally, the bias and all feature loss functions are fused using a weighted approach, and the foreign object features are used as comparison values to output the foreign object recognition results. The experimental data shows that the lowest loss function of this method is only 0.16, and there is no significant fluctuation with the increase of the number of training samples. This method can accurately identify foreign objects on coal mine belt conveyors without missing or misidentifying them, and the recognition speed does not exceed 0.8 seconds. The above results indicate that the method can accurately, efficiently, and stably identify foreign objects.