Research Of Object Detection Method Based On Shallow Feature Fusion And Channel Selection

Object detection is a long-term basic problem in the field of computer vision,and it has been an active research field for decades.With the birth of new computer equipment,there seems to be a new breakthrough in the field of object detection recently.Many scholars began to look back to the past and found a new direction for the development of object detection technology in the field of machine learning.With the concept of deep learning on the stage of the new era,object detection technology has also entered the era of convolutional neural networks from the era of manual features in the past.In recent years,object detection algorithms based on deep learning have sprung up like mushrooms after a rain.According to the different design concepts of models,they can be roughly divided into two-stage and single-stage.Their respective characteristics are also very distinct.The two-stage model is relatively high in detection accuracy,while the single-stage model is relatively fast in detection speed.Based on the single-stage object detection algorithm,this thesis designs an object detection method based on shallow feature fusion and adaptive channel selection.Firstly,a shallow feature fusion structure is proposed,which based on the classical single-stage object detection network.In this structure,deep separable convolution and anti-aliased convolution are used to carry out subsampling feature fusion for shallow feature maps.According to the different times of subsampling,four types of subsampling feature fusion modules are designed respectively to realize deep mining and utilization of the information contained in shallow feature maps.Secondly,an adaptive channel selection process is designed.In this process,the feature maps after feature fusion are firstly combined,and then the weight of each channel is redistributed,so as to improve the contribution of effective feature maps to model parameter updating.Finally,a loss function based on cosine distance is designed,which can improve the accuracy of object detection algorithm when classifying targets by using the properties of cosine distance.Experiments show that the proposed method is more effective on public datasets than many existing algorithms.In order that the improved object detection algorithm can be applied to mobile devices,this thesis selects the lightweight improvement on the basis of the above method.Firstly,the backbone network is replaced with a lightweight network which has a smaller number of parameters.Then,in order to avoid the problem of insufficient feature extraction caused by the simple backbone network,the receptive field module is proposed to enrich the feature extraction results of the network.Finally,in view of the imbalance of positive and negative samples generated by the single-stage object detection algorithm in the detection process,a gradient harmonic cross-entropy loss function is proposed to balance the contribution of positive and negative samples in the updating of model parameters.The experiment shows that although the lightweight improvement of the model will bring a certain degree of precision loss,the detection speed will be improved quite significant and can achieve real-time detection on mobile devices.