LME SSD: Light-Multi Extract modules based Single Shot Detector for Resource-Restricted Usages.

For a long time, the accuracy of the object detection methods has been continuously improved, but the computational cost and resource occupation have not received enough attention. In recent years, with the development of some excellent lightweight object detection methods for compressing model parameters and computational quantities, the object detection framework can run on resource-constrained mobile devices. This also makes reducing computational cost and resource occupation an important direction to improve the performance of the object detection methods. In this study, we have proposed an efficient detection framework named LME (Light-Multi Extract) SSD, which maintains detection accuracy while reducing the occupation of computing resources. We propose the LME module, which compresses parameters and calculations of the architecture by applying depth-wise convolutions with different convolution kernel sizes to extract features. In addition, we have further improved the accuracy of the detection framework through LF (Light-Fusion) module. LF performs bidirectional fusion of feature map information by up-sampling of bilinear interpolation and down-sampling of depth-wise convolution, thereby improving the accuracy while maintaining the scale of the detection framework. We follow the SSD method to predict by setting anchor boxes on feature maps of different sizes.