Object Detection Based On Attention Mechanism

Object detection is one of the most challenging research tasks in the field of computer vision.It requires the computer to classify the objects of interest in an image containing multi-categories,and locate each object by using the bounding box in the image.In practical applications,object detection has very important research significance and value in intelligent monitoring,space remote sensing and medical assistant diagnosis.With the improvement of GPU computing performance and the continuous study of artificial neural network,current object detection algorithms based on deep learning have made remarkable achievements in generic object detection tasks.However,the performance improvement of small object recognition is still unsatisfactory.Besides,unlike the detection of generic objects of public datasets,for some specific scenarios in reality,such as commodity object detection,due to the characteristics of background environment and object distribution,there are also specific difficulties of recognition,where there is still much room for improvement.Therefore,on the basis of the research of the current popular detection algorithms,this dissertation studies the detection on the general objects and the commodity objects respectively,including the following two aspects:Currently,the detection task on benchmark dataset MSCOCO has acquired great performance.However,due to the large variety and variable size of the objects in the dataset,the overall recognition performance still has a large space to improve,especially the recognition accuracy of small-sized objects.To this end,this dissertation proposes a detection algorithm framework based on dual attention model,specifically including the Multi-scale Feature Cascade Attention(MFCA)module and the Salient Channel Self-Attention(SCSA)module,which makes the feature map affected by multiple attention to become more representative.Comparing with the state-of-the-art algorithms,extensive experiments verify the effectiveness of the method.With the upgrade of consumption,intelligent supermarket and warehouse management are becoming increasingly important.In order to study the task of commodity object detection,this dissertation introduce a small and dense commodity dataset with milk tea on the shelf,and employs the current mainstream detection methods to evaluate the dataset.Aiming at the recognition difficulties of small,dense and large average number of objects,a multi-scale receptive field feature fusion attention model is proposed.The model constructs multi-scale feature pyramid through bottom-up feedforward neural network,and then carries out fusion operation in different scale features from top to bottom to enrich the semantic information of features,which enhances the information of the positives in the feature map and reduces the interference of background information and negative objects by weighing each area of the feature map of convolution neural network with different importance.In addition,a new loss function is proposed,which can assign the weights of loss contribution automatically to makes the problem of category imbalance improved effectively,thus acquiring a superior detection performance.