L-SSD: lightweight SSD target detection based on depth-separable convolution

The current target detection algorithm based on deep learning has many redundant convolution calculations, which are difficult to apply to low-energy mobile devices, such as intelligent inspection robots and automatic driving. To solve this problem, we propose a lightweight target detection algorithm, L-SSD, based on depth-separable convolution. First, we chose the lightweight network MobileNetv2 as the backbone feature extraction network, and we proposed an upsampling feature fusion module (UFFM) to fuse the output feature maps of MobileNetv2. Deep semantic information is introduced into the shallow feature map to improve the feature extraction capability while reducing the complexity of the model. Second, we propose a local-global feature extraction module (LGFEM), which uses LGFEM to generate five additional feature layers to expand the feature map's receptive field and improve the model's detection accuracy. Then, we use an improved weighted bidirectional feature pyramid (BiFPN) for feature fusion to construct a new feature pyramid that fully utilizes the feature information between different layers. Finally, we propose asymmetric spatial attention (ASA) that enhances the expression ability of the features before BiFPN feature fusion, providing good positional information for the feature pyramid. Experimental results on the PASCAL VOC and MS COCO datasets show that the model parameters and model complexity of L-SSD are reduced by 85.9% and 96.1%, respectively, compared to SSD. A detection speed of 106 frames per second was achieved in NVIDIA GeForce RTX 3060 with detection accuracies of 73.8% and 22.4%, respectively. The optimal balance of model parameters, model complexity, detection accuracy, and speed are achieved.