Multi-scene small object detection with modified YOLOv4

Abstract

Listen

Translate article

Applications of small object detection exist in many different scenarios of our daily life, and this topic is also one of the most difficult issues in studying object detection and recognition. Therefore, improving small object detection accuracy is not only important in theory, but also significant in practice. However, current detection relevant algorithms are inefficient in this task; therefore in this study, a type of generalized improved algorithm based on YOLOv4 model is proposed. After the “ADD” and “Concat” layers of the conventional cross stage partial network (CSPNet), a type of mixed attention module (MA) is added to strengthen the spatial and channel feature information of the object; the dilated convolution module (DC) is then used to expand the receptive field information around the peripheries of target small objects. Several numerical experiments show that this modified model associated with its algorithms achieves higher mean average precision of the tests on the Pascal VOC and the VISDRONE data sets (mAP@0.5). Compared with YOLOv4, these performance assessment indices increase by 3.23% and 2.88% respectively. After refining the YOLOv4 network structure with the condition that the size of input images set as 416*416, the average accuracies of the model can reach 87.58% on Pascal VOC data set and 26.12% on VISDRONE data set. These results present the apparent better performance of this method than the original detection algorithm on recognizing small objects without affecting the real-time performance.