Research On Lightweight Target Detection Algorithm Combining Edge Information

With the rapid development of artificial intelligence and deep learning,object detection technology has become one of the current research hotspots,and has broad application prospects in many fields such as autonomous driving,remote sensing,video surveillance,and robot vision.The traditional target detection algorithm is not only difficult to design,but also cannot fully extract the feature information of the target.The detection accuracy of the algorithm depends on the hand-designed feature representation and feature extraction algorithm,which also leads to the unsatisfactory object detection effect.At present,the detection accuracy of target detection algorithms based on deep learning far exceeds that of traditional algorithms and has become the mainstream.However,the current convolutional neural network model has high hardware requirements,and the algorithm with higher accuracy tends to have a larger model,and the larger the model,the slower the detection speed,which leads to the lack of real-time performance of the target detection algorithm with high accuracy.With the development of lightweight convolutional neural networks,lightweight object detection models have also received more and more attention.In order to improve the detection speed of the target detection algorithm,meet the real-time requirements,and realize the lightweight of the target detection algorithm,the paper studies the FCOS algorithm,related lightweight methods and improved model methods.First of all,FCOS is a single-stage anchor-free target detection model,which even exceeds some two-stage and anchor-base detection models in detection accuracy.However,there is still a problem that the detection speed is slow and cannot be detected in real time.Aiming at the problem that the reasoning speed of FCOS algorithm cannot meet the requirement of real-time detection,a lightweight object detection model called ILFCOS is proposed.The ILFCOS model introduces the Mobile Nete V2 structure as the backbone network,and reduces the size of the model while improving the model detection speed by replacing the ordinary convolution with a depthwise separable convolution.Then,the method of reducing the size of the input,pruning redundant branches of the feature pyramid network,and multi-scale training greatly improves the detection speed while maintaining the detection accuracy.Next,the edge information extraction module is integrated,and the directional gradient extracted by the Sobel operator is added to increase the dimension of the input data and enhance the edge semantic information.The Mish activation function is used to replace the Re LU activation function in the head part of the model to retain the semantic information of the negative part and improve the generalization ability of the model.Using DIo U as the evaluation index,it can more accurately describe the similarity between the predicted frame and the target frame,and optimize the regression loss value of the target frame.The model algorithm is optimized from the above three aspects,and the final experimental results on the MS-COCO2014 dataset show that the average accuracy of the ILFCOS algorithm reaches 34.0,and the number of images detected per second reaches 72.5,which can meet the needs of real-time detection.Finally,for the problem of low detection accuracy of the ILFCOS algorithm,an attention mechanism is introduced.An improved CA module is proposed,along with two methods of fusing attention mechanisms.By incorporating the improved CA module into the backbone network and feature pyramid,the capabilities of model feature extraction and feature fusion are improved.The final experimental results on the MS-COCO2014 dataset show that the average accuracy of the ILFCOS algorithm incorporating the attention mechanism reaches 35.8,and the number of images detected per second reaches 62.3.Compared with the ILFCOS algorithm itself,although the detection speed drops by 10.2FPS,it still far exceeds the demand for real-time detection.Compared with the original FCOS algorithm,with a difference of only 2.8 in accuracy,the detection speed is increased by about three times,and the optimization of the model is successfully completed.