Abstract: Instrument detection is an indispensable part of intelligent instrument testing, its effect directly determines the accuracy of instrument testing. In view of the complex positioning background of the instrument and the requirement of fast detection speed, a target detection algorithm based on improved YOLOv3 is proposed. Based on YOLOv3 algorithm, the last two network blocks in the Darknet are first replaced with DenseNet (Densely Connected Convolutional Networks) so as to enhance the reuse of features by the model. And then the lightweight Darknet-48 is used as feature extraction networks, and the convolution neural network in the DenseNet is modified to the deep separable convolution network, and then  the 6 layer convolution before all detection layers (YOLO Detection) is modified to 2 layers so as to reduce the parameters of the model. At the same time, GDIOU bounding box is introduced to regress coordinates loss, and  the weight of the loss function is readjusted according to the detection requirements. Experimental results show that compared with the original algorithm, the number of parameters of the improved YOLOv3 algorithm is reduced by 40%, and the accuracy  and recall  in instrument detection reach 95.83% and 94.98%, respectively, which is increased by 2.21 percentage points and 2.09 percentage points. The average accuracy is increased by 2.42 percentage points  and the detection speed is increased by 30.18%.

Key words: instrument detection, lightweight, DenseNet, depth separable convolution, loss function