Small Target Detection Optimization Algorithm Based on Improved YOLOv5

Abstract

The accuracy of the traditional model algorithms mainly oriented to detection of small target is generally low, the unstable problem, we have made modifications on the basis of YOLOv5 and proposed a more suitable algorithm for detecting small targets. First of all, increased the effective channel attention mechanism, solves the convolution the characteristics of the loss, as a result of the global context information loss in the process of problem, improves the fuzzy image cases of small target detection accuracy. Secondly, the network structure is reconstructed, the multi-scale network model is established, and the target detection layer is combined to improve the detection performance of remote sensing small target model at different scales under the feature information extraction network. Finally, by integrating the loss function SIOU into the whole model, the improved algorithm has made progress in many aspects. The improved algorithm is compared with the original YOLOv5 model. The data obtained from the experiment mean that on the VOC data sets, the accuracy value of ours is higher than original algorithm 0.7%, 0.8% percentage points higher than YOLOv5s small target detection accuracy.