SlimYOLOv4: lightweight object detector based on YOLOv4

Object detection is a valuable but challenging technology in computer vision research. Although existing methods could attain satisfactory results on high-performance computers, but the huge number of network parameters brings great operating pressure to the mobile devices with limited computing power. Existing methods are usually in a dilemma between accuracy and speed. The low detection effect brings great difficulties to the implementation of detection tasks. This paper optimizes the classic YOLOv4 and proposes the SlimYOLOv4 network structure. Firstly, we change the feature extraction network from CSPDarknet53 to MobileNetV2. Secondly, more appropriate DO-DConv (depthwise over-parameterized depthwise convolutional layer) and DSC (depthwise separable convolution) were selected to replace the standard convolution in the network structure, which greatly reduces computation and improves network performance. Finally, Leaky ReLU is replaced by ReLU6 to improve the numerical resolution. We evaluate SlimYOLOv4 on Pascal VOC07+12 dataset and MS COCO dataset. The experimental results demonstrate that the parameters of our method account for only 12.6 % of YOLOv4, and the speed is 1.59 times that of YOLOv4, reaching 60.19 frames per second (FPS), which is suitable for real-time detection. It achieve 70.83 % mean average precision (mAP) on PASCAL VOC07+12 and 29.2 % mAP on the MS COCO dataset. As a lightweight object detector, it takes into account both speed and accuracy, which can be comparable to the state-of-the-art detectors as well.