Abstract: In the long-term transportation process, the mining aramid belt conveyance equipment will produce scratch, smash and other damages.The surface defects of aramid strip should detect timely, but the traditional machine vision detection has some problems such as low accuracy, large background interference, high rate of missing detection and false detection.You only look once unified real-time object detection (YOLO) is proposed to use deep learning neural network detection.In the field test, YOLOV3 neural network has a low identification accuracy and insufficient sensitivity to small targets.In this paper, YOLOV3 neural network is optimized, and the transmission process of network information is replaced by DenseNet (dense connection network) with more complete feature expression from the previous ResNet (residual network).Meanwhile, convolution dimension reduction is applied for optimization to reduce detection time.After comparison in the field, the optimized YOLOV3 algorithm improved the detection accuracy by 26% compared with the previous one through frequency domain transformation and Otsu algorithm.Compared with the non-optimized YOLOV3 algorithm, the detection accuracy is increased by 15%.Through field experiments, this method could effectively improve the detection of defects on the small target of aryl ribbon.