Research On Household Garbage Image Detection And Classification Method Based On Deep Learnin

Speed and accurate is of high importance in the application of intelligent domestic trash image detection.Effective trash detection algorithms can improve the economic efficiency of trash sorting as well as adapt to more end devices with low arithmetic power.Since the scenarios in which domestic trash appears in daily life scenarios are often accompanied by complexity and diversity,we need to apply advanced algorithms that are suitable for detecting multiple trash objects in complex scenarios.However,many existing domestic trash detection algorithms suffer from insufficient detection accuracy and the common problem of excessive amount of detection algorithm parameters.To accurately detect and classify domestic trash in these different scenes,we first need to prepare rich trash image datasets and effective data enhancement methods,and then train algorithmic models that are sensitive to multi-scale objects and complex image backgrounds.This paper will combine advanced deep learning techniques and dataset optimization methods to address the existing challenges in domestic trash image detection and classification.The main tasks are as follows:(1)In response to the lack of existing domestic trash image datasets,this thesis proposes a self-made domestic trash dataset containing 10,000 target samples of domestic trash with 17 trash category samples and standardized labeling.In addition,various data augmentation research works are conducted for the dataset problems of too few samples of some categories,single scenario and uneven distribution.The training effect of the data is enhanced by the dataset enhancement method,and the most suitable data enhancement strategy is obtained by data enhancement experiments.(2)A new involution channel convolution operator is applied for channel feature extraction.The traditional CNN two-dimensional convolutional spatial feature extraction method is optimized from the perspective of channel convolution.The spatial convolution feature extraction method is replaced by the channel convolution feature extraction method,and the channel convolution group is divided by the sequence processing idea.The iCSPLayer feature extraction layer is constructed by using involution to improve the accuracy of domestic trash detection and classification while reducing the number of neural network parameters and increasing the detection speed.(3)Based on the one-stage object detection algorithm — YOLOX,i-YOLOX object detection algorithm is proposed for domestic trash detection and classification.Combining CBAM feature attention mechanism and channel convolution mechanism,iPANet structure is proposed for feature enhancement extraction of the algorithm,and iResHead residual decoupling detection head structure is proposed.By improving these algorithmic structures,the network is able to learn richer semantic and texture features during training,which effectively improves the detection model’s accuracy in identifying garbage targets in real-life scenarios.This thesis selects YOLOX-S as the baseline algorithm and makes multiple improvements.We evaluated multiple accuracy indicators and network parameters,and found that in the algorithm improvement experiment,compared to the baseline algorithm i-YOLOX,the average precision increased by 1.47%,the number of parameters decreased by 23.3%,and the detection speed increased by 40.4%.In addition,we conducted extension experiments on the i-YOLOX algorithm proposed in this study and performed cross-comparison experiments using the publicly available dataset Trash Net.The experimental results showed that the i-YOLOX algorithm performed well in comprehensive indicators.