Research And Application Of Complex Object Detection Method Based On Lightweight Object Detector

With the continuous application of deep learning in object detection,related research in the field of object detection has made rapid progress.The detection accuracy for general objects is already very high,reaching over 80% on public datasets such as Pascal VOC.However,detectors often struggle with complex objects such as small targets and low-light targets,resulting in low detection rates or even the inability to detect them.Therefore,this paper focuses on these two complex targets,low-light targets and small targets,and proposes two detection algorithms to effectively improve the detection accuracy for these types of targets.To tackle the challenges associated with detecting targets in low-light conditions,an algorithm for low-light object detection is introduced,which combines image enhancement techniques with network lightweighting.The algorithm aims to overcome the difficulties in extracting features from low-light targets.In order to achieve this,the low-light images are first subjected to enhancement using an enhanced version of Mobile Netv1 and Gaussian process,which enhances the feature extraction capability of the low-light images.To optimize the feature extraction,deep separable convolutions are employed,along with a refinement of the feature enhancement network that reduces unnecessary connections and parameters.This results in a significant reduction in computational complexity while maintaining high accuracy.Furthermore,the anchor box selection process adopts the anchor-free approach,which further reduces the computational parameters required.However,to address any potential decrease in accuracy associated with using anchor-free,the sim OTA method is improved to achieve a better balance between positive and negative samples,thereby enhancing the training speed of the model.The effectiveness of the proposed method is validated through experiments,which demonstrate a noteworthy 2.81% improvement in detection accuracy on the Ex Dark dataset for low-light images.To address the challenges in detecting small targets,a target detection algorithm based on contextual features and attention mechanism is proposed,which improves the YOLOv3 network.The number of feature extraction layers is increased,and the network structure is deepened to enhance the capability of extracting features from small targets.Different connection methods are employed for different network layers to better detect targets.For different datasets,adjusting the size of the anchor boxes poses certain challenges when using the k-means method for anchor box selection.Therefore,to better address the issue of adapting the value of k to the size of anchor boxes,this paper proposes combining the manual adjustment of anchor boxes in SSD with k-means.This approach is used to match different quantities of anchor boxes and adjust their sizes accordingly.Contextual feature fusion and attention mechanism are utilized to address the issue of information loss for small targets after multiple convolutions.The modified model can better detect small targets,achieving a 5.71%improvement in detection accuracy on the dataset of the 2020 Underwater Target Detection Algorithm Competition.