Research On Small And Weak Object Detection In Images Based On Feature Enhancement

Object detection aims to classify and locate objects of interest in images,which is widely used in practical scenarios such as face recognition,license plate recognition,and autonomous driving.However,due to small scale,weak feature response and the pooling operation of the network which is easy to lose small object feature,small and weak object detection in complex background is a difficult task.In order to solve the above problems,the researchers propose the strategy based on multi-scale features,using high-resolution feature maps to retain the feature of small and weak objects,and then use feature enhancement methods to improve the representation ability of small and weak object feature.However,there are still the following problems:(1)There is a problem of strong interference of background features in shallow feature.Due to the lack of prior knowledge of background,it is difficult for the network to depict background feature,resulting in background feature interfering with the feature represen-tation of small and weak objects.(2)The progressive feature fusion method leads to the problem of gradual fading of the semantic information from deep feature in the transmis-sion process.How to effectively realize the direct interaction between deep feature and shallow feature is the key to solving the problem.(1)In terms of object feature description,aiming at the strong interference problem of background feature in shallow feature,thesis proposes an attention-based local feature layer enhancement method.Specifically,thesis designs an attention module combined with singular value decomposition,which uses singular value decomposition to map the feature map to hidden space and filters the background information with the singular value as the threshold.And then thesis adaptively extracts the object feature by using the atten-tion mechanism of the multi-receptive field.The proposed method improves by 1.5%compared with the AP_sindicator of the baseline method.(2)In terms of feature fusion of multi-level features,thesis proposes a global feature layer enhancement method based on context information,aiming at the problem that the semantic information of deep feature is gradually weakened during transmission.Specif-ically,thesis designs a global information interaction module,which extracts the context information of each level of feature maps through the method of spatial information ag-gregation.And then the module uses multilayer perceptrons to realize the interaction of context information.Finally,the module uses context information to guide the feature maps of each level,further promoting the transmission of semantic information in the fea-ture maps of each level.The proposed method is improved by 1.4%compared with the AP_sindicator of the baseline method.(3)In terms of network design,in order to improve the detection accuracy of small and weak objects,thesis constructs a small and weak object detection network based on feature enhancement.Specifically,a global-local feature layer enhancement network is designed,which combines the object feature enhancement of shallow feature and the direct interaction of cross-layer feature to improve the feature representation ability of small and weak objects.At the same time,this paper uses the higher-resolution feature map to retain more feature information of small and weak objects,and then adds high-resolution detection branch,which effectively improves the detection accuracy of the network for small and weak objects.Compared with the AP indicator of the existing method,the proposed method is improved by 0.8%.