BFANet: Bidirectional feature aggregation network for efficient and accurate object detection

Recently, As artificial intelligence and computer vision technology develops, object detection applications have become increasingly widespread. Efficient object detectors have powerful capabilities to accurately and efficiently analyze video clips captured by cameras, providing important decision-making basis for people. Existing real-time detectors are relatively well defined on large targets, but suffer from insufficient detection accuracy on small and medium-sized targets, such as occluded detection targets, and ambiguous detection objects. Based on the existing Yolov7 detection model, super-upsampling and super-dowansampling methods are designed in this paper. In the super-downsampling section, a bidirectional feature aggregation model was designed to change the ordinary convolution. In the super-upsampling section, the snake convolution and carafe module are used instead of the ordinary convolution. In addition, to address the issue of decreasing recall as accuracy increases, the SPD module has been introduced, which can improve recall without decreasing accuracy. we perform experiments on two public datasets (VOC and Visdrone) for verify the result of the new network in the article and compared with some other state-of-the-art methods. Our Experimental results prove that we proposed model for target detection achieves promising results. On the VOC dataset, the new network improved the object detection accuracy and recall by $3.1 \%$ and $2 \%$, respectively. And on the VisDrone dataset, the proposed network achieved an AP and AP50 of $36.8 \%$ and $\mathbf{6 0. 6 \%}$ for object detection, respectively.