Waste Detection Based On Mask R-CNN

In recent years, with the rapid development of the world economy and the continuous improvement of people's consumption levels, a large amount of waste has been generated. The random discarding, simple stacking, and subsequent treatment of this waste are causing many problems. For example, it destroys the ecological environment, pollutes water, soil, and air, leads to a large number of mosquitoes and bacteria, and increases the probability of infectious diseases. The implementation of waste classification can effectively improve the living environment and reduce waste pollution to the environment, which is conducive to ensuring people's health and sustainable economic development. It can also help people classify and recycle waste more effectively, so it improves the efficiency of waste recycling. Based on deep learning algorithms and object detection technology, this thesis implements the detection and classification of waste and focuses on the following three goals: 1) Use VIA to perform annotation and classification. 2) Mask R-CNN algorithm is then used to deal with the detection of waste; 3) By applying mask shapes of different sizes, this thesis tests and analyzes the performance of relevant models applied to waste detection. This thesis selects six categories of waste as the standard for detection and classification. The six categories are the following: paper, plastic, glass, metal, trash, and compost. There are 2451 images in the dataset, with a size of 1024 × 768 in JPG format. In the data processing stage, VIA image annotation tool is used to manually annotate all images and generate independent JSON structure files. This step includes the classification of the waste category, border, and mask. According to the format of COCO dataset, all independent JSON files are merged into one JSON file, which is composed of images, categories, and annotations, and contains all image information. The dataset was divided into 10 pieces, of which 8 are used as training datasets and 2 as validation datasets. Mask R-CNN is used to perform waste detection. It is a classic instance segmentation model with a two-stage framework. In the first stage, it traverses the image and generates proposal regions that may contain waste. Resnet101 is used as feature extraction network. In the second stage, it extracts the precise location and range of waste according to the proposal regions, this stage has 3 branch tasks: classification, bounding box regression, and segmentation. The first 2 branches are used to obtain the waste category and the position of the bounding box. The segmentation branch is used to generate the mask of trash on the pixel level. To evaluate the results, this thesis uses mAP (mean average precision) as the evaluation index of trash detection. Different mask shape sizes are tested to evaluate the performance of waste detection in order to find the best one. The experimental results show that this algorithm can accomplish such a task with good performance. In particular, the best performing Mask R-CNN is able to classify each category with a precision of up to 82.85%. This was achieved by using a mask shape of 28*28 pixels. When the light is good, the object detection results are better; in a few cases, white plastics are classified as paper because most of the shapes and colors of paper and plastics are similar with only local features. To conclude, this paper verifies the feasibility and effectiveness of the proposed method in the waste detection task by obtaining valuable results.