CAL-SSD: lightweight SSD object detection based on coordinated attention

Although existing object detection algorithms have achieved excellent detection accuracy, with the continuous improvement of detection accuracy, the parameters of the model are getting larger and larger, and the model complexity is getting higher and higher, which makes it difficult to deploy the object detection algorithms on the edge end and mobile end. In order to improve the application of the object detection algorithm on edge and mobile, this paper proposes a lightweight object detection algorithm, CAL-SSD, using a coordinated attention mechanism. First, we embed the coordinated attention mechanism into MobileNetv2 to form CA_MobileNetv2 as the backbone of the CAL-SSD object detection algorithm, significantly reducing the model parameters and complexity and improving the network’s ability to differentiate between object and background. Second, we design a super-resolution feature fusion module (SFFM) to introduce deep semantic information into shallow feature maps. Then, we use depthwise separable convolution instead of traditional 3×3 convolution to construct additional feature layers and detection heads, further reducing the model parameters. Finally, we employ BiFPN to construct a new feature pyramid to utilize the multi-scale features of the target fully. Experimental results on the PASCAL VOC and MS COCO datasets show that CAL-SSD significantly reduces the model parameters and complexity and achieves an optimal balance of speed and accuracy.