Research On Small Object Detection Algorithm Based On Feature Fusion And Attention Residual Network

Object Detection is a significant component of computer vision technology.It serves as a foundation for downstream visual tasks and is applied extensively in various detection scenes with intelligence,such as self-driving,satellite remote sensing images,and industrial inspection.In general scenarios,the presently object detection technologies can basically meet the production demands of detection performance for the objects of large-medium.However,for small size objects,the detection effect is always not ideal.Therefore,small object detection has become a bottleneck that restricts the further development of object detection in this field.Due to its small size,the performance of small object detection is low,which is also the main reason.As a result,it occupies a small amount of pixels in the image,the features are not obvious,and it is prone to interference from environmental noise.Therefore,it is difficult to perform feature extraction on small objects through the network.YOLO-v7 is a model of a one-stage object detection algorithm based on deep learning,which can effectively balance the accuracy and speed of object detection.However,due to the characteristics of small objects mentioned above,the effectiveness of this algorithm in small object detection is not ideal.This article conducts research on this issue,and the main work is as follows:Firstly,due to the small object’s small pixel proportion in the image and its key features are easy to lose,which weakens its feature problem in the multi-task learning structure of the feature fusion pyramid,the Left Fusion Factor PANet（LFPA）feature fusion structure based on L-αfusion factor is proposed.By using the number of objects on adjacent layers to determine the L-αfusion factor and adding the L-αfusion factor into the inter-layer structure of the feature pyramid,the information transfer ratio between adjacent layers can be controlled,so as to solve the problem that small objects have insufficient feature expression ability under the influence of the inter-layer information game.On the one hand,achieving small objects can gain more influence of backpropagation parameters in multitasking learning,and on the other hand,strengthening the importance of small object features in websites to further improve the network’s ability to detect small objects.Aiming at the top-down and bottom-up feature fusion pathways of the path aggregation network PANet in YOLO-v7,the optimal location for adding fusion factors was proposed and experiments were conducted to verify that L-αfusion factor is the most conducive to improving the information imbalance in the process of feature fusion,and the improved model can achieve the best effect of small object detection.Secondly,the structure of Detach and Merge（DM）is proposed.In order to reduce or even disappear the limited feature information of small objects due to the information loss caused by feature extraction based on network subsampling,the MP module of YOLO-v7 is improved and the sampling is completed through the operation of detaching and merging feature maps.Thus,the information loss of small object in the process of feature extraction can be reduced and its feature expression ability can be enhanced.Propose the Grouped Residual（GR）structure to address the issues of low resolution and difficulty in extracting feature information from small objects.By using group residual convolution to extract small object features based on different channel dimensions,the model’s ability to perceive small object features is enhanced.The DM structure is composed of DM and GR.Multi-Scale Squeeze Attention（MSA）is proposed to solve the problem of fuzzy edge of small objects and easy interference by environmental noise,so that the network can establish long range channel dependence,and the reuse of feature information of small objects can be enhanced through the use of residual network.Has reached the network characteristics of small object key attention to improve,and weaken the irrelevant background information.The experiment proves that the proposed DM and MSA modules play an effective role in improving the small object detection performance of YOLO-v7.In summary,this study proposes an improved YOLO-v7 algorithm based on feature fusion and attention residual networks.Compared with the original YOLO-v7 algorithm,the improved algorithm improves the₅₀⁴9))by 3.89%and the₅₀⁸（67）7))by 3.11%on the Tiny Person dataset.And in MS COCO dataset,increases by 2.5%andincreases by4.3%,which has surpassed the current mainstream algorithms,thus proving that the algorithm verified by the research is effective and advanced for small object detection performance.