The separation of coal and gangue is a crucial step in coal production. Traditional methods, such as manual identification, mechanical methods, and X-ray transmission, are labor-intensive and environmentally harmful. Image-based coal and gangue recognition technology, known for its high intelligence, compact equipment, and eco-friendliness, has become a research hotspot in dry coal beneficiation. However, scholars have found that camera-based image acquisition of coal and gangue is suboptimal under certain conditions, such as in dim environments or with adhesive samples, and the recognition effectiveness is unstable under varying lighting. To address these issues, we propose a method for coal and gangue feature extraction and recognition based on laser speckle imaging.

We analyze the characteristics of laser speckle and design a coal and gangue laser speckle imaging system. A dataset of laser speckle images under varying illuminance is constructed for experimental validation. A region-of-interest extraction method is developed to retain the target area of coal and gangue laser speckles under different lighting conditions while minimizing edge interference, thereby obtaining more accurate feature data. Feature extraction methods are designed to better capture the intra-class similarity and inter-class dissimilarity of minerals. A support vector machine (SVM) is employed to recognize coal and gangue, verifying the method’s effectiveness. The feature vector extracted from the gray-level size zone matrix (GLSZM) is input into the SVM to validate its effectiveness. We compare the recognition performance of the SVM when using the fusion of gray-level features and gray-level co-occurrence matrix (GLCM) features with that of using the fusion of these two features plus GLSZM features. This confirms the enhancement in recognition effectiveness of coal and gangue by our method. The recognition accuracy of our method is compared with prevalent coal and gangue recognition methods under various illuminance conditions, verifying its effectiveness, particularly in low-light and fluctuating illuminance environments.

The GLSZM feature alone demonstrates an accuracy rate of 91.7% (Table 1), indicating its effectiveness in recognizing coal and gangue laser speckle images. Compared to the commonly used fusion of gray-level and GLCM features, the multi-dimensional feature recognition accuracy, recall rate, and precision rate of our method’s fusion of GLSZM features improve by 2.8%, 2.7%, and 2.8%, respectively. Our laser speckle imaging-based method significantly improves recognition accuracy compared to natural light methods (Fig. 9), with a maximum increase of 18.0%. Across six different illuminance levels, the average accuracy of the laser speckle recognition method is 96.05%, with a standard deviation of 1.85%, while the natural light method achieves an average accuracy of 81.25%, with a standard deviation of 3.90%. These results demonstrate that our method effectively improves recognition accuracy and exhibits more stability under varying lighting conditions.

We propose a method for feature extraction and recognition of coal and gangue based on laser speckle imaging. By applying a Gaussian pyramid in the Lab color space and using the Otsu threshold for image segmentation, we effectively preserve the target speckle areas while reducing edge interference under varying lighting conditions. Additionally, we construct a method to extract regions of interest, yielding more accurate feature data. A texture feature extraction method based on the GLSZM is proposed, revealing pronounced intra-class similarity and inter-class differences between coal and gangue. By combining the GLCM and gray histogram, we extract both gray-level and texture features, establishing a multi-dimensional feature extraction method. The SVM classifier, trained on these features, improves recognition accuracy by an average of 14.8% across different lighting conditions, with the highest improvement of 18.0%. The standard deviation of the recognition accuracy rate is reduced from 3.90% to 1.85%, indicating that our method is less affected by lighting variations and offers more reliable and stable recognition under complex lighting environments.