Research Of Target Fast Detection Based On Deep Learning

In the field of computer vision,target detection includes two main steps: feature extraction of image information,target location and recognition.Machine learning algorithm is a commonly used theory in target detection.The Deep Learning(DL),as a hot area of machine learning,possesses a strong ability to generate deep feature information,therefore becomes valuable method to detect the target.The target detection method based on DL has excellent performance but consumes huge computing resources,which means the detection method is hard to implement and deploy in low latency scenarios.Thus,target fast detection has important research significance.As for DL-based target detection,two kinds of target fast detection models are proposed: one is based on hybrid structure CNN and the other is based on feature multiplexed CNN.The main works of the paper are as follows:(1)The Faster R-CNN model and CNN model are mainly studied in the DL-based target detection.From the perspective of the network structure,the impact of these two models on the detection speed is analyzed.After that,data set with special needs is created and used for follow-up work.Through some experiments,the performance of these models is evaluated.(2)A new target fast detection model based on hybrid structure CNN is proposed.Based on the typical CNN model,a multi-structure hybrid CNN model is designed,and this model is integrated into the Faster R-CNN model to achieve the target fast detection model.The experimental results based on self-built data sets illustrate that the target fast detection model based on hybrid structure CNN will reduce the model parameters,improve the real-time performance of the model,and it also alleviate the memory consumption of the model.(3)A new target fast detection model based on feature multiplexed CNN is proposed.The hierarchical features of each convolutional layer in the CNN model are fused,which is effective for improving the performance of the CNN model.Combined with feature de-scaling technology,a feature multiplexed strategy is proposed.The strategy is applied to the hybrid structure CNN model and combined with Faster R-CNN model to achieve the target fast detection model.The experimental results illustrate that the feature multiplexed strategy will improve the detection accuracy and have a good improvement on the detection of small targets.