| Object detection,one of the most critical components of computer vision technology,aims to accurately estimate the coordinates of the target in the input image and video and try to identify the categories of all targets correctly.A convolutional neural network is applied to perform feature extraction for pictures for the most advanced computer vision system.However,due to the strong learning ability of the convolutional neural network,it may easily lead to a problem of overfitting.Model over-fitting leads to a decline in small occluded objects’ classification or detection performance.Specifically,the target detection effect decreases sharply and increases the target occlusion degree and the scale difference between the targets.To this end,this paper takes the general data set as the primary research object.It proposes a multi-scale detection algorithm for occluded things based on a convolutional neural network.Besides,related experimental studies are carried out.In summary,it requires urgent solutions for occlusion problems and multiple scales for the current object detection.On this basis,this paper aims to provide guidance for image classification and accessible object detection from occlusion and various scales.In this paper,a detailed analysis is conducted on the problems of multiple scales and occlusion.Regarding the method,the convolutional neural network is adopted to develop a general object detection system based on deep learning.As an object detection system,such a method is characterized by unprecedented robustness and accuracy.This paper targets designing a general object detection method,through which it is expected to promote the development of available object detection.This paper mainly tries to study issues as follows.A data enhancement algorithm is proposed based on corner points.Based on the corner points detector of an improved local autocorrelation function,this algorithm aims to detect the corner points of the images in the training image set,thus obtaining high-quality corner points.Then on the original image,different corner point-centered local areas with high quality are partially occluded to simulate partial occlusion in various positions,which leads to a new training image set.Secondly,the expanded training set is applied to the convolutional neural network.In this way,the algorithm innovatively incorporates traditional vision technology and convolutional neural network,which shows a good fitting and effective feature extraction performance.In summary,such a method strengthens the algorithm’s ability in occluded processing targets while reducing the false positive rate and probability of miss in the case of occlusion,thereby improving the model’s performance.An improved residual network is proposed to capture high-quality image features.Based on the improved residual network,this method attempts to weight and fuse the feature maps before and after down-sampling,which leads to a newly generated down-sampled feature map.Secondly,it is transmitted to the output layer,through which it can obtain better learning target characteristics and naturally weaken the influence of multi-scale targets on target detection.Compared with the current mainstream,namely,the occlusion processing algorithm,such a way adopts the down-sampling method of the attention mechanism,whose effective feature extraction plays a critical role in enhancing the performance of object detection and target classification.This way aims to reduce the false positive rate and probability of miss under the multi-scale situation.According to the results of comparative experiments,some conclusions can be drawn as follows.In this paper,a data enhancement algorithm is proposed based on corner points.Compared with the previous data enhancement methods,such an algorithm shows different performance in the test set.The results show that such a method demonstrates a higher average accuracy.In terms of backbone network construction,the experimental results indicate the highest accuracy rate for this network.The network performs outstandingly regarding issues of multiple scales and occluded targets.In summary,this study offers a potential solution to the long-existing problems: occlusion and numerous scales of target detection. |
