The accurate recognition of mixed materials on conveyor belts in the feeding process of blast furnaces (BFs) is of great significance for the reasonable distribution of raw materials and ensures stable production of BFs. However, material images usually suffer from complex backgrounds, varying particle sizes, and material stacking, which make it challenging to detect mixed materials. Although the existing two-stage object detection methods have high accuracy, they have some shortcomings in detecting dense and small edge objects. To this end, a mixed material recognition method based on object detection is proposed to convert the problem of detecting mixed materials into the detection and quantification of coke. First, an object detection method based on composite feature reweighting is proposed for coke. A composite feature extraction (CFE) module is designed to express more complex semantic characteristics. A serial attention mechanism module is also introduced to enhance the network’s attention to the region of interest. Finally, to accurately count the pixels occupied by the coke, the segmentation method is used to segment each detected coke block. The experimental results show that the proposed method performs well on our constructed material image dataset, reaching 94.48% on mAP.