Research On Small Target Detection Algorithm In Complex Scenes Based On Neural Network

Target detection algorithms are widely used in tasks such as traffic control,video surveillance,unmanned driving,and medical image recognition.The detection speed is slow,and at the same time,the feature extraction and fusion part does not pay enough attention to small targets,and it is difficult for the effective small target information to be finally received by the detector.Therefore,the detection of small targets becomes a detection problem,and the average accuracy is always no more than 25%.In this paper,the detection accuracy of small targets in complex scenes is not high,and the real-time performance of the detection system is poor.The main research work and research results are as follows:1)A multi-scale switchable atrous convolution SPC based on feature pyramid is proposed.This module is combined with the Neck part of the YOLOv4 baseline network,and a 3-D convolution is constructed in the feature pyramid,and the shared convolution check with different amplitude changes is used.To detect the same target in different pyramid levels,this module optimizes the convolution layer,extracts features from the same image through convolution kernels of different sizes,and then uses the Switch function to select and integrate the extraction results,effectively expanding the convolution kernel.It realizes repeated observation of target features at multiple scales,and effectively solves the problems of difficult feature extraction and low detection accuracy for small target detection in complex scenes.Then,the improved focal loss loss function is applied to further solve the problem of the number of sample classifications and the unbalanced sample distribution in the network model.The method proposed in this paper performs well on the MS COCO dataset,improving the average accuracy by 5.4% on the basis of Yolov4,reaching 48.9%,and reaching 29.6 FPS in 608×608 images.2)After completing the target detection task,the system is further improved to perform more subtle image instance segmentation tasks,the YOLOv4 network is improved based on the U-net network framework,and the mask branch is added to the head part for instance segmentation tasks.The scale attention module is combined with the YOLOv4 neck part to improve the network’s instance segmentation accuracy for small objects.A composite loss function for instance segmentation is designed to solve the problem of sample imbalance and network overfitting at the same time.The multi-scale attentionYOLOv4 network proposed in this paper achieves an average segmentation accuracy of39.7% at 24.6 FPS on the MS COCO dataset,and the detection result for small objects reaches 22.5%,which is 2.0% higher than the baseline network.