Research on small target detection methods based on deep learning

At present, object detection based on deep learning has achieved very good results. Due to the difficulties of small target detection, such as small target, low resolution and complex background, there are still problems such as low precision, high missed detection rate and low real-time performance in detection. In view of the above problems, in order to effectively improve the performance of small target detection, the following improvements are made on the basis of the existing target detection algorithm, and the work is as follows: Based on the advantages of the one-stage detection algorithm YOLOv5, which has high real-time performance and high precision in small target detection, a small target detection algorithm based on regional context learning and feature enhancement is proposed. Aiming at the problem of small targets in small target detection, hollow convolution of multiple convolution rates is used to obtain different receptive fields, and then the feature information of the target and its adjacent regions is obtained. The information of adjacent regions plays an important complementary role in small target detection. It can effectively solve the problem of small targets: for the interference problem with complex background, feature enhancement module is introduced in the backbone of the detection algorithm to further improve the feature extraction capability of the backbone, and the attention mechanism is combined with the modules in the original backbone to make the network more able to focus on small target information and avoid the interference of complex background. In the above algorithms, the location information extraction and prediction frame screening algorithms are further optimized, and a small target detection algorithm based on multi-scale fusion and improved non-maximum suppression algorithm is proposed. In the multi-scale fusion part of the detection algorithm, this algorithm combines the shallow feature layer with the deep detection layer to increase the accuracy of target positioning and boundary regression by taking into account both shallow position information and deep semantic information, and is more conducive to small target detection. In the screening and detection box of the detection algorithm, the center point distance and aspect ratio penalty items are added to the non-maximum algorithm. In this way, in the screening process of the detection box, not only the confidence and intersection ratio are considered, which can effectively reduce the missed detection rate of small targets. Finally, the improved algorithm is tested on TT100K and VOC, and good results are obtained in the precision and recall rate of small targets.