The accurate detection of coke overflow in high-dust environments is a pivotal challenge in achieving intelligent coke loading. A method was proposed to address this issue for the intelligent detection of coke loading overflow, which was based on dark channel prior knowledge and the ResNet network. Firstly, a video collector was used to obtain video information of the coke loading scene, and the original time-series video image frames were processed to obtain the region of interest between the discharge port and the loader. Secondly, the prior knowledge method of dark channels was employed to process the regions of interest. Enhancing the contrast between the target areas and irrelevant areas within the regions of interest, thereby mitigating the effects of dust on subsequent detection models. Moreover, the problem of overflow detection was transformed into a binary classification task by labeling the regions of interest based on the actual loading of coke. Finally, the ResNet network was utilized for modeling, enabling the completion of model training and experimentation during the loading process of newly acquired coke. The experimental results demonstrate that the proposed method exhibits promising performance on new data, achieving an overall accuracy of 86.81%. Specifically, the accuracy, recall, and F1 score for the overflow class are 84.12%, 90.74%, and 0.873 0, respectively. Furthermore, the application of the dark channel prior algorithm in data processing results in a notable increase in the recall rate of the overflow class by 3.31%.