Intelligent monitoring small target detection algorithm based on improved faster R-CNN

Abstract

Current applications of the small target detection algorithm include enterprise logo detection, small face detection, pedestrian detection, traffic sign detection, automatic driving, remote sensing image detection, and criminal investigation. Practical applications are primarily still in the research phase. Consequently, this paper proposes an optimized small target detection algorithm on the basis of Faster R-CNN. Utilizing ResNet-18 with fewer convolutional layers in place of ResNet-50 as the backbone network of the original Faster R-CNN algorithm, the feature extraction network is strengthened, reducing the number of parameters of the algorithm. Moreover, by analyzing the characteristics of small targets, we can eliminate missed detection and false detection of small targets, Faster R-CNN algorithm uses feature fusion to amplify the contrast between small target features and background features in the feature pyramid network; Enhance prediction box deviations caused by pooling regions of interest by utilizing ROI Align. Lastly, the improved algorithm is experimentally verified on the MSCOCO data set. The faster R-CNN small target detection algorithm proposed in this thesis improves the optimized algorithm’s average accuracy from 89.14% to 93.39%, which is 4.25% higher than the original Faster R-CNN algorithm. Intelligent monitoring significantly enhances the detection and recognition of smaller and medium-sized targets.