Research On Small Object Detection Method Based On Feature Enhancement

Small object detection is widely used in the fields of automatic driving,medical lesion detection,military guidance,and industrial defect detection.Compared with medium and large objects,small objects are more difficult to be detected due to the fact that small objects have fewer pixel points,contain less feature information,and are susceptible to interference from environmental factors such as illumination occlusion.The existing small object detection algorithms are basically researched under the framework of object detection,but the current methods are ineffective in detecting small objects because the features of the small objects are easy to lose and difficult to extract after down-sampling.To address the problems of difficult extraction of small object feature information and insufficient shallow contextual semantic information,small object feature enhancement is carried out from the aspects of feature extraction and feature fusion under the existing basic framework of object detection to improve its detection performance.The specific research work is as follows:(1)Aiming at the problem that the features of small objects are easy to lose after downsampling,which makes it difficult to extract the feature information of small objects,a small object detection algorithm based on multi-scale feature extraction is proposed.This method improves the problem of poor detection of small objects by improving the feature extraction capability of the backbone network.Firstly,the backbone enhanced network I-Darknet53 is proposed under the basic framework of SSD.The network improves the network structure of Darknet-53 by constructing a new grouping residual structure,which effectively improves the feature channel extraction capability of the backbone network.Then a multi-scale feature enhancement module composed of parallel multi-branch dilated convolution is used to enhance the multi-level feature maps extracted by the backbone enhanced network to strengthen the network’s association with multi-scale feature information.Finally,the loss function of the network is optimized to effectively alleviate the sample imbalance problem during the training process by controlling the weight of positive and negative samples.The proposed small object detection algorithm is tested on the public datasets.The experimental results show that the proposed method can effectively improve the detection performance of small objects.(2)To solve the problem of insufficient shallow contextual semantic information,a small object detection algorithm based on multi-level feature fusion is proposed.This method enhances the semantic information of shallow feature maps through multi-level feature fusion of deep and shallow feature maps.Firstly,the different levels of feature maps generated by the feature extraction network are used to form a feature pyramid for object detection.Then a multilevel feature fusion strategy is used to perform deep feature fusion between shallow and deep layers to obtain shallow enhanced semantic features.This multi-level fusion strategy is based on the basic feature fusion by adding an efficient channel attention guided module consisting of an attention mechanism,which directs the different feature channels to be weighted,and make the network pay more attention to the important small object semantic feature information,finally enabling more efficient feature fusion between shallow and deep layers.The experimental results on the public datasets validate that the proposed method can further enhance the network’s ability to discriminate small objects.(3)Given that the actual production process of industrial products,due to the failure of mechanical equipment and improper manual operation,small defects may appear on the surface of the product,and the detection of these small object defects by the workers will be affected by subjective factors,resulting in miss detection and false detection of surface defects.To solve this problem and combine the proposed small object detection algorithm,a Web-based small object surface defect detection system platform is built,and each functional module is developed in detail.The testing results show that the system can meet the basic requirements for real-time detection of small object surface defects in industrial scenarios.