Abstract: Currently, some intelligent devices are available to assist in the detection of imperfect wheat grains. However, the background of grain surface images acquired by intelligent devices is dense and complicated with overlapping particles, causing noise interferences in the detection of imperfect wheat grains. To address the high missed detection rate of imperfect grains in target detection algorithms and to enhance the model detection speed, this study optimized the lightweight network model YOLOV4-Tiny. First, a small target detection layer was added to enhance the utilization of high semantic information. Then, the SENet attention mechanism optimized with exponential thinking was embedded to facilitated the design of an Enhanced Feature Extraction Network (Increase-FPN) in order to enhance the model’s ability to extract features of imperfect grains amidst complex backgrounds so that the detection accuracy could be improved and false negative rates reduced. At last, depthwise separable convolution was employed as the feature extraction method for the residual network of the backbone component to reduce the calculation of model parameters, optimize model deployment, and solve the issue of poor real-time performance. Experimental results demonstrated that the improved IDS-YOLO algorithm achieved a balance between detection speed and accuracy, with an average increase of 6.2% in mean average precision (mAP) when compared with other benchmark algorithms. The frames per second (FPS) value was 88.03, meeting the real-time detection requirements, and the parameter size of the improved model was only 5.51 MB.

Key words: real-time detection; imperfect wheat grains; small target detection; grain storage quality; deep learning